People ask us the same questions about AI search every week, and the industry argues about most of them. Here is what we see in our own data: the answers AI platforms gave about the brands we measure, the bots that visit our website, and the visits AI platforms send. Each answer says what we measured, what we found, what to do, and what we don't know yet.
Measuring what AI says about your brand
Should branded and non-branded prompts be reported together?
No. Blending them produces a mention rate that describes neither. In our data, the blend reported a brand in 39.0% of answers, while category prompts mentioned it in 11.3%.
What we measured. Atlas classifies each tracked prompt as branded (it includes the brand's name), comparison (it compares the brand with a named competitor) or category (it names no brand). We scored 129,652 answers for 21 brands, September 1 to October 18, 2026.
What we found. Answers to branded prompts mentioned the brand 85.8% of the time (25,643 answers). Answers to comparison prompts mentioned it 75.5% of the time (26,140 answers). Answers to category prompts mentioned it 11.3% of the time (77,869 answers). All prompts blended together produced 39.0%.
129,652 answers · 21 brands · September 1 to October 18, 2026
What it means for you. An answer to a prompt that includes your name will mention you almost every time, so counting those mentions tells you little. For category prompts, measure how often answers mention and recommend you. For branded prompts, measure how much of the cited material is yours and how often the answer warns buyers about you (see How often do AI answers warn buyers about a brand?). Ask any vendor or report for these numbers separately.
What we don't know yet. The blended figure depends on how many prompts of each kind a brand tracks, so another mix would give another blend. That dependence is the problem with blending.
Related research: How to choose defensible prompt targets, Branded vs discovery prompts in Atlas.
Last checked October 19, 2026.
Can I trust one run of a prompt?
Not for a single prompt. When a brand appeared in an AI answer to a category prompt, a third of the time another answer to the same prompt that day left it out.
What we measured. Atlas asks the same prompt on the same platform more than once on the same day. We call that a repeat group. We looked at category prompts from July 13 to August 31, 2026. A repeat group averaged 2.9 answers. We kept the 1,039 groups where at least one answer mentioned the brand.
What we found. In 359 of those 1,039 groups (34.6%), at least one answer mentioned the brand and at least one did not. On ChatGPT, Claude and Gemini without web search, it happened in 51.5% of groups. On chat platforms with web search, including Perplexity, it happened in 49.3%. On Google AI Overviews, Google AI Mode, Google's organic results and Microsoft Copilot, it happened in 8.1%.
1,039 repeat groups · July 13 to August 31, 2026 · 11 platforms
What it means for you. One run of one prompt on a chat platform is close to a coin flip for whether your brand appears, when it appears at all. Read results across many prompts and repeated runs, and report rates with their sample sizes. Google's AI features are steadier, so a single reading there is more useful.
What we don't know yet. With about three answers per group, we undercount disagreement. More runs per group would find more. Across all prompts, including those where the brand never appears, mention agreement is 95.0%, because a brand that is never mentioned is "stable." We report the conditional figure because it is the one a buyer feels.
Last checked October 19, 2026.
How many prompts do I need to track?
For a rate on one platform, about 100 prompts gets the margin of error to roughly ±5 points. Running the same prompts again helps far less than adding prompts.
What we measured. We took 68,982 AI answers to prompts that do not include the brand's name, for 21 brands on 10 AI platforms, from September 1 to October 18, 2026. For each brand and platform we resampled prompts 1,000 times to see how much the mention rate moves with 10, 25, 50, 100 or 200 prompts. We used 5,982 groups of same-day repeat answers from July and August to measure how much a second run of the same prompt adds.
What we found. At a mention rate near 10%, the margin of error on one platform was ±15 points with 10 prompts, ±8 with 50, ±5.5 with 100 and ±4 with 200. On chat platforms, 76% of the variation came from which prompt was asked and 24% from asking it again. On Google AI Overviews and AI Mode, 98% came from the prompt. Three runs of 100 prompts were about as precise as 119 prompts run once on chat platforms, and 101 on Google. Week-to-week changes in a brand's mention rate were no larger than chance in 260 of 265 comparisons.
68,982 answers · 21 brands · September 1 to October 18, 2026 · 10 platforms
What it means for you. Put the budget into more prompts before more repeats. Keep the prompt set fixed between readings, so a change reflects the answers and not a different set of prompts. Read a change of 2 or 3 points on one platform as noise unless it continues across several readings. A single prompt is different: on chat platforms, its answer can include the brand on one run and leave it out on the next (see Can I trust one run of a prompt?), so repeats matter when you judge one prompt.
What we don't know yet. The repeat measurements come from summer data for 10 brands. They do not include changes in what the web says from one week to the next. We have not tested brands whose mention rates sit near 50%, where margins are widest.
Related research: Options for tracking AI visibility and performance.
Last checked October 19, 2026.
What is the difference between a mention, a citation and a recommendation?
They are three separate events, and they often do not happen together. On category prompts, AI answers cited a brand's website without mentioning the brand more often than they did both.
What we measured. For each scored answer, Atlas records whether it mentions the brand, whether it cites a page on the brand's own website and whether it recommends the brand. We looked at 53,200 answers with sources to category prompts for 21 brands, September 1 to October 18, 2026.
What we found. 5,246 answers (9.9%) cited the brand's website and did not mention the brand. 4,340 (8.2%) cited the website and mentioned the brand. 2,395 (4.5%) mentioned the brand without citing its website. Across all 77,869 scored category answers, 11.3% mentioned the brand and 3.6% recommended it.
53,200 answers with sources · 21 brands · September 1 to October 18, 2026
What it means for you. A citation means the platform used your page as a source, and in our data that often happened while the answer said nothing about you. A mention means buyers saw your name. A recommendation means the answer told them to pick you. Report all three, and treat the recommendation rate as the closest to a business outcome.
What we don't know yet. We count a citation only for the brand's main domain. Pages on other domains a brand controls are missed, so the citation counts are a floor.
Related research: Introducing Atlas by Momentic.
Last checked October 19, 2026.
If AI cites my website, is it more likely to recommend me?
Only a little, once you account for being mentioned. Answers that cited a brand's website mentioned the brand far more often. Among answers that mentioned the brand, citing its website changed the recommendation rate by 2.7 points.
What we measured. We compared category prompt answers that cited the brand's own website with those that did not. The window is September 1 to October 18, 2026, across 21 brands. Then we repeated the comparison among answers that mentioned the brand.
What we found. Answers that cited the brand's website mentioned the brand 45.3% of the time and recommended it 15.2% of the time (9,586 answers). Answers that did not cite it mentioned the brand 5.5% of the time and recommended it 1.7% of the time (43,614 answers). Among answers that mentioned the brand, the recommendation rate was 33.7% with a citation of its website (4,340 answers) and 31.0% without one (2,395 answers). Seven brands had at least 30 answers on each side. For 4 of them the citation went with a higher recommendation rate, and for 3 with a lower one.
53,200 answers with sources · 21 brands · September 1 to October 18, 2026
What it means for you. Most of the raw gap comes from being mentioned at all. A citation of your website goes along with a mention, and once you are mentioned, the citation adds little to the recommendation rate in our data. Getting into the answer is the larger step. What third parties say about you likely does more of the work after that.
What we don't know yet. This is an association. A platform may cite your website because it already decided to mention you (see Does ChatGPT search specific websites when it answers?). We have not run a controlled test of adding or removing pages.
Related research: Introducing Atlas by Momentic.
Last checked October 19, 2026.
How often do AI answers warn buyers about a brand?
About one answer in three, when the buyer names the brand. When the prompt names no brand, warnings about the brand were rare.
What we measured. Atlas flags an answer when it includes a caveat about the brand. A caveat is a warning or qualification, such as complaints about billing or a note that a product lacks a feature. We scored 129,652 answers for 21 brands, September 1 to October 18, 2026.
What we found. Answers to branded prompts included a caveat 33.6% of the time (25,643 answers). Answers to comparison prompts included one 21.9% of the time (26,140 answers). Answers to category prompts included one 1.8% of the time (77,869 answers).
129,652 answers · 21 brands · September 1 to October 18, 2026
What it means for you. A buyer who already knows your name and asks about you will often hear a "but." Read the caveats in answers to your branded prompts, find which sources they cite, and fix the facts that are wrong or out of date. That is reputation work as much as search work.
What we don't know yet. The caveat rate depends on how the prompts are worded (see Does the wording of a prompt change the answer?). Our classifier decides what counts as a caveat, and it has not been audited against human labels in this window.
Related research: How to choose defensible prompt targets.
Last checked October 19, 2026.
Does the wording of a prompt change the answer?
Yes, a lot. Branded prompts with negative wording drew a caveat three times as often as neutral ones, and the negative tone carried into the searches platforms ran.
What we measured. Atlas labels each tracked prompt's framing as negative ("Is [brand] a scam?"), neutral or positive. We compared caveat rates for branded prompts, September 1 to October 18, 2026, across 21 brands. Separately, in September 2026 we classified the searches ChatGPT, Gemini and Claude ran while answering (query fan out searches) by the framing of the prompt that started them.
What we found. Answers to negatively worded branded prompts included a caveat 56.4% of the time (2,395 answers). Answers to neutral branded prompts included one 18.7% of the time (14,981 answers). Positively worded branded prompts drew a caveat 40.9% of the time (1,126 answers). In the query fan out data, 41.98% of searches started by negative prompts were negative themselves (908 of 2,163). For neutral prompts it was 0.90% (400 of 44,204).
25,643 branded answers · 21 brands · September 1 to October 18, 2026
What it means for you. A prompt set with more negative prompts reports more warnings, whatever the brand does. Balance the framing of tracked prompts, and report negative prompts separately. When you compare brands, compare them on prompt sets with the same framing balance.
What we don't know yet. Positive prompts drawing more caveats than neutral ones is unexplained. That sample is small (1,126 answers). We have not tested why. The query fan out framing figures come from one month.
Related research: How to choose defensible prompt targets.
Last checked October 19, 2026.
Is there an AI ranking position for my brand?
Not a stable one. On chat platforms, repeated answers to the same prompt on the same day listed the same set of brands only 60% to 74% of the time. An average position across those lists hides that.
What we measured. For same-day repeat groups (the same prompt on the same platform on the same day, two or more answers), we checked whether every answer listed exactly the same set of brands. The window is July 13 to August 18, 2026. It covers 26,712 groups across 10 platforms and 15 brands.
What we found. Google AI Overviews listed an identical set in 91.0% of groups, and Google AI Mode in 88.0%. Microsoft Copilot was at 84.5%. ChatGPT without web search was at 73.6%. Perplexity was at 64.9%, Claude with web search at 62.6%, Gemini with web search at 60.8% and Gemini without web search at 60.0%.
26,712 repeat groups · 15 brands · July 13 to August 18, 2026 · 10 platforms
What it means for you. If the set of brands changes this often, position within the set changes more. Track how often answers mention and recommend you across many runs. A recommendation rate holds more meaning than an average rank.
What we don't know yet. Our identical-set test treats order as part of the set, so it reads conservative. Groups where no brand appeared count as identical, which inflates these rates. Among groups that named any company, the identical-set rate was 48.6%.
Last checked October 19, 2026.
Can an AI citation be tied to revenue?
Not one citation to one sale. We have not found a defensible way to do that, and we do not claim it. Direction and size over time can be measured.
What we measured. Atlas measures answers (mentions, citations, recommendations) and, where GA4 is connected, AI referral sessions. It does not join a specific answer to a specific visit. No platform tells a website which answer a visitor read.
What we found. For 11 brands with GA4 connected, AI referral sessions were 11,692 from September 1 to October 18, 2026. That count includes only the visits that kept their referrer (see How do I track AI traffic in GA4?). AI answers also change visits that come later through branded search or Direct, and GA4 cannot see that link.
11,692 AI referral sessions · 11 brands · September 1 to October 18, 2026
What it means for you. Report AI results as rates (how often answers mention and recommend you) next to business trends (branded search, Direct, AI referral sessions, leads). Judge direction and size together over months. Be wary of any report that assigns revenue to individual AI citations.
What we don't know yet. We have not measured how much branded search or Direct traffic AI answers cause. A study we have planned would start on that.
Related research: Options for tracking AI visibility and performance.
Last checked October 19, 2026.
How do I track AI traffic in GA4?
Use a custom channel group that matches AI assistant referrers by name, and treat the result as a floor. Some AI visits arrive with no referrer and land in Direct.
What we measured. Atlas reads AI referral sessions for 11 brands with GA4 connected (see Which AI platforms send the most visitors?). We have not measured how many AI visits arrive without a referrer.
What we found. Across the 11 brands, ChatGPT sent 94.0% of the AI referral sessions GA4 could identify, September 1 to October 18, 2026. In our client work, we have seen consent banners set up in a way that makes ordinary visits look like a flood of Direct traffic. That pattern can hide or mimic AI traffic.
11,692 AI referral sessions · 11 brands · September 1 to October 18, 2026
What it means for you. Build a channel group with referrer rules for chatgpt.com, gemini.google.com, perplexity.ai, copilot.microsoft.com, claude.ai and similar domains, and put it above the default rules. Read key events by event name, because a default "conversions" total can count different actions for different properties. Then compare AI referral sessions against the trend in Direct, not against a guess.
What we don't know yet. We do not know the share of AI visits that lose their referrer. A study we have planned would estimate it by comparing referral sessions with landing pages that only AI answers link to.
Related research: Options for tracking AI visibility and performance.
Last checked October 19, 2026.
Which AI platforms matter
Which AI platforms send the most visitors?
ChatGPT, by a wide margin. It sent 94.0% of the AI referral sessions we measured across 11 brands.
What we measured. Atlas reads Google Analytics 4 sessions that arrive from AI assistants for each brand with a connected GA4 property. The window is September 1 to October 18, 2026, across 11 brands. It counts earned visits only and excludes paid traffic.
What we found. The 11 brands received 11,692 AI referral sessions. ChatGPT sent 10,989 (94.0%). Gemini sent 437 (3.7%). Perplexity sent 117 (1.0%). Microsoft Copilot sent 76 (0.7%). Claude sent 34 (0.3%). For the median brand, ChatGPT's share was 93.1%. The lowest single brand was 46.2%.
11,692 AI referral sessions · 11 brands · September 1 to October 18, 2026
What it means for you. If you can only look closely at one platform's traffic, look at ChatGPT. Do not read referral share as audience share, though. People use other assistants heavily and click out less, or arrive without a referrer.
What we don't know yet. GA4 records a visit as AI referral only when the referrer survives. Visits from apps often arrive with no referrer and show up as Direct. One brand provided 43.1% of all sessions, so the total leans toward that brand, and that is why we also give the median.
Related research: Top AI chatbots and assistants by market share.
Last checked October 19, 2026.
Do AI platforms cite the same kinds of sources?
No. Each platform has its own source habits, and the differences are large enough to change your plan by platform.
What we measured. Atlas recorded every source cited in scored answers on 8 platforms for 21 brands, September 1 to October 18, 2026. That is 760,722 citations. We counted the share going to four large websites.
What we found. Reddit was 7.55% of Google's organic results for our prompts and 0.00% of Claude's citations. YouTube was 5.37% of Google AI Overviews citations and 0.01% of ChatGPT's. Wikipedia was 3.13% of Claude's citations and under 1% everywhere else. Across all platforms, Reddit was 2.16% of citations, LinkedIn 1.09%, YouTube 1.31% and Wikipedia 0.96%.
760,722 citations · 21 brands · September 1 to October 18, 2026 · 8 platforms
What it means for you. "Get on Reddit" or "make videos" is advice for some platforms and not others. Pick the platforms your buyers use, then look at which sources those platforms cite for your prompts.
What we don't know yet. Shares depend on the industries in our sample. We have not tested why a platform avoids a source. Claude's zero for Reddit could reflect access limits, ranking choices or both.
Last checked October 19, 2026.
Do AI platforms cite the same sources?
Mostly not. Two AI platforms answering the same prompt at the same time share about one in ten cited websites. One platform asked the same prompt a week later shares about four in ten with its earlier answer.
What we measured. Atlas asks the same prompts to 7 AI platforms in the same run. We compared the websites each platform cited for the same prompt, for 21 brands from September 1 to October 18, 2026. Overlap is the number of websites both platforms cited, divided by the number either one cited. We compared 125,545 pairs of answers to prompts that do not include the brand's name.
What we found. Two different platforms had an average overlap of 0.097 on the same prompt in the same run. The same platform had an average overlap of 0.410 with its own answer to that prompt a week later. That gap held for all 17 brands we could test. When all 7 platforms cited sources for a prompt, 74.2% of the cited websites appeared in one platform's answer only, and 0.12% appeared in all 7. Google AI Overviews and Google AI Mode overlapped most with each other and with Google's organic results. ChatGPT with web search overlapped least with the others.
125,545 answer pairs · 21 brands · September 1 to October 18, 2026 · 7 platforms
What it means for you. A source plan built from one platform's citations covers a small part of what the other platforms cite. Pull cited sources per platform for your prompts, and plan for the platforms your buyers use (see Do AI platforms cite the same kinds of sources? and Which AI platforms send the most visitors?). The few websites that several platforms cite for your category are the ones to check first.
What we don't know yet. We have not tested why platforms differ. Their search indexes, how many sources each one shows, and how each one ranks sources are all possible reasons. The prompts are the ones we track for 21 brands, so levels may differ in other industries.
Related research: Top AI chatbots and assistants by market share.
Last checked October 19, 2026.
Do AI answers always include sources?
No. Platforms that search the web cited sources in most answers. The same models without web search cited none.
What we measured. Atlas recorded whether each scored answer included at least one source. The window is July 13 to August 15, 2026, across 11 platforms.
What we found. Microsoft Copilot and Google's organic results included sources 100.0% of the time. Perplexity and Gemini with web search included them 99.2% of the time. Claude with web search was at 98.9%, Google AI Mode at 95.9% and Google AI Overviews at 78.5%. ChatGPT with web search was at 88.3% after we fixed a defect in our own collection. ChatGPT, Claude and Gemini without web search included sources 0.0% of the time.
Scored answers · July 13 to August 15, 2026 · 11 platforms
What it means for you. When an AI platform answers from memory, no page of yours can be cited in that answer. What the model learned in training decides what it says. Measure answers with and without web search separately (see Should I measure ChatGPT with and without web search?).
What we don't know yet. We measured the API versions of these platforms with default settings, on desktop, in one locale. The consumer apps decide on their own when to search, and we cannot see that decision directly.
Last checked October 19, 2026.
Should I measure ChatGPT with and without web search?
Yes. On category prompts, the web search versions of ChatGPT, Claude and Gemini mentioned the measured brand more often than the same platforms answering from memory.
What we measured. Atlas asks the same category prompts to ChatGPT, Claude and Gemini with web search and without it. The window is September 1 to October 18, 2026, across 21 brands. Each lane had about 7,000 scored answers.
What we found. ChatGPT mentioned the brand in 13.2% of answers with web search and 8.9% without. Claude mentioned it in 15.8% with web search and 6.8% without. Gemini mentioned it in 16.0% with web search and 10.8% without.
About 7,000 scored answers per lane · 21 brands · September 1 to October 18, 2026 · 3 platforms with and without web search
What it means for you. The answer from memory reflects what the model learned in training. The answer with web search reflects what the platform finds and reads today. Pages and third-party coverage you publish now can change the second number within weeks. The first changes when a new model version is trained. Track both, and expect different timelines.
What we don't know yet. The consumer apps mix both modes, and we cannot see which mode a given person gets. The models in the two lanes can differ for some platforms, which adds a second variable.
Last checked October 19, 2026.
How AI platforms search
What searches do AI platforms run behind an answer, and how many?
Between one and four per answer on the platforms that show them. ChatGPT and Gemini ran about 3.4. Claude ran about 1.5.
What we measured. Query fan out is the set of searches an AI platform runs while it writes an answer. Three platforms report these searches to us: ChatGPT with web search, Gemini with web search and Claude with web search. We collected them for 21 brands from September 1 to October 18, 2026.
What we found. ChatGPT ran 39,810 searches across 11,727 answers, 3.39 per answer. Gemini ran 39,428 across 11,701 answers, 3.37 per answer. Claude ran 17,317 across 11,328 answers, 1.53 per answer. The searches were short: 7.5 words on average for ChatGPT, 5.0 for Gemini and 6.8 for Claude.
96,555 searches · 21 brands · September 1 to October 18, 2026 · 3 platforms
What it means for you. An AI answer to one prompt draws on several searches, and those searches read like short keyword queries. Pages that answer those searches directly are the ones a platform can find. Look at the actual query fan out searches for your prompts and check whether a page of yours answers each one.
What we don't know yet. Google AI Mode, AI Overviews, Copilot and Perplexity do not report their searches, so we cannot see them. Searches vary from run to run. In a 10-run test of one platform in August, no search string appeared in all 10 runs.
Related research: What Verbatim search reveals about query fan-out.
Last checked October 19, 2026.
Does ChatGPT search specific websites when it answers?
Yes, often. ChatGPT ran at least one site-restricted search in 44.4% of its answers. Claude and Gemini almost never did.
What we measured. A site-restricted search limits results to one website, written as "site:example.com." We counted these among ChatGPT's query fan out searches for 21 brands, September 1 to October 18, 2026.
What we found. ChatGPT ran 16,105 site-restricted searches, 40.5% of its 39,810 searches. 5,212 of its 11,727 answers (44.4%) included at least one. The searches targeted 2,131 different websites. 2,672 (16.6%) targeted the measured brand's own website. 1,924 (11.9%) targeted government or university websites. 277 (1.7%) targeted large platforms such as Reddit, YouTube and Wikipedia. Claude ran site-restricted searches in 0.05% of its searches, and Gemini in none.
39,810 searches · 11,727 answers · 21 brands · September 1 to October 18, 2026
What it means for you. ChatGPT often goes to a company's own website to check a fact before it answers. Make sure the facts it is likely to look up (pricing, locations, eligibility, policies, product specifics) are on your website, current and easy to find in the page text.
What we don't know yet. Most site-restricted searches went to websites we did not classify, including other brands and publishers. The share rose sharply in late August, so this behavior can change with model updates.
Related research: What Verbatim search reveals about query fan-out.
Last checked October 19, 2026.
Do AI platforms add the year to their searches?
Claude does, often. A third of Claude's searches included a year. ChatGPT added one in 8.5% of its searches and Gemini in 1.3%.
What we measured. We checked each query fan out search for the years 2025 or 2026. The window is September 1 to October 18, 2026, across 21 brands.
What we found. 33.7% of Claude's 17,317 searches included 2025 or 2026, and 26.9% included 2026. 8.5% of ChatGPT's 39,810 searches included one of those years, and 7.2% included 2026. 1.3% of Gemini's 39,428 searches included one, and 0.9% included 2026.
96,555 searches · 21 brands · September 1 to October 18, 2026 · 3 platforms
What it means for you. When a platform adds "2026" to a search, pages that show the current year in the title, heading or text match that search more closely. Keep the year on pages where it is honest, such as annual reviews and current pricing, and update the content when you update the date.
What we don't know yet. We have not measured whether pages that show the current year get cited more often (see Does updating content or dates raise AI citations?). We did not separate prompts that already contained a year from those that did not.
Last checked October 19, 2026.
Do Google rankings predict which pages AI Overviews and AI Mode cite?
Partly. One in five pages AI Overviews cited was in Google's top 5 organic results for the same prompt that day. For AI Mode it was about one in seven.
What we measured. Atlas sends each tracked prompt to Google's organic results, AI Overviews and AI Mode on the same day. It stores the top 5 organic results. We compared the pages AI Overviews and AI Mode cited with those top 5 results for the same prompt on the same day, September 1 to October 18, 2026, across 21 brands.
What we found. 20.4% of the 63,235 pages AI Overviews cited were in the top 5 organic results (9,062 prompt-days). At the website level, 35.2% came from a website in the top 5. For AI Mode, 13.7% of 45,713 cited pages were in the top 5, and 30.3% came from a website in the top 5 (10,174 prompt-days).
108,948 cited pages · 21 brands · September 1 to October 18, 2026 · 3 Google surfaces
What it means for you. Ranking in the top 5 helps, and most of what Google's AI features cite comes from outside it. AI Mode leans less on the top results than AI Overviews. Keep doing SEO, and also look at which pages Google's AI features cite for your prompts, because many of them will not be the ones you see on page one.
What we don't know yet. We compare with the top 5 only, so pages ranked 6 to 10 count as outside, and the true overlap with page one is higher. Our prompts are conversational, so they may rank differently from the short keywords people type into Google.
Related research: Re: GEO, AIO, AEO, SEO.
Last checked October 19, 2026.
Does AI search replace SEO?
No. In our data, the same work feeds both. Google's AI features cite pages that rank, and AI platforms run short keyword searches before they answer.
What we measured. We did not run a single study for this question. The answer draws on three measurements in this FAQ: Google's AI features citing top 5 pages (see Do Google rankings predict which pages AI Overviews and AI Mode cite?), the searches AI platforms run (see What searches do AI platforms run behind an answer, and how many?) and the visits AI platforms send (see Which AI platforms send the most visitors?).
What we found. 20.4% of pages AI Overviews cited ranked in Google's top 5 for the same prompt. ChatGPT and Gemini ran about 3.4 short searches per answer. AI referral sessions across 11 brands were 11,692 over seven weeks.
Draws on the measurements for Google rankings, query fan out and AI referral sessions
What it means for you. Keep the SEO basics: crawlable pages, clear titles, pages that answer real searches, and links from respected websites. Add the work AI answers reward on top: facts on your website that platforms can check, and coverage on the third-party websites they cite. Budget for both from one plan.
What we don't know yet. We have not measured AI referral sessions as a share of total organic sessions across the portfolio. A study we have planned would.
Related research: Re: GEO, AIO, AEO, SEO.
Last checked October 19, 2026.
Where AI answers get their sources
How much of what AI cites about my category is my own website?
Very little. On category prompts, a brand's own website was 3.0% of what AI platforms cited.
What we measured. We counted citations to each brand's own domain in AI answers for 21 brands, September 1 to October 18, 2026. We excluded Google's organic results and counted only AI platforms. Prompts were grouped by type (see Should branded and non-branded prompts be reported together?).
What we found. On category prompts, the brand's website was 3.0% of 410,143 citations. On comparison prompts it was 12.6% of 164,798. On branded prompts it was 31.6% of 119,525.
694,466 citations · 21 brands · September 1 to October 18, 2026 · AI platforms only
What it means for you. When a buyer asks about your category without naming you, AI platforms build the answer mostly from other websites. Your own pages still count. They are what platforms check when they look you up (see Does ChatGPT search specific websites when it answers?). Plan for both: accurate, complete pages on your website, and accurate coverage on the third-party websites cited for your category.
What we don't know yet. We count only the brand's main domain, so subdomains on other domains and profiles a brand controls on other websites are not counted as owned. The true owned share is somewhat higher.
Last checked October 19, 2026.
How much do AI answers cite Reddit?
Reddit was a small share of citations and appeared in a larger share of answers. Across AI platforms, it was 1.7% of citations and appeared in 11.6% of answers with sources. It depends heavily on the platform.
What we measured. Atlas recorded every source cited in scored answers on 7 AI platforms for 21 brands, September 1 to October 18, 2026. We counted Reddit's share of citations and the share of answers that cited Reddit at least once.
What we found. Reddit was 1.72% of 703,273 citations. It appeared in 11.6% of 78,802 answers with sources. By platform, Reddit appeared in 20.0% of Google AI Overviews answers, 17.1% of Google AI Mode answers, 16.1% of Gemini answers, 15.7% of Microsoft Copilot answers and 12.7% of Perplexity answers. It appeared in 1.0% of ChatGPT answers and in none of Claude's. By brand, Reddit's share of citations ranged from 0.13% to 4.18%, with a median of 1.51%. YouTube followed a similar pattern. It appeared in 23.2% of AI Overviews answers and in almost no ChatGPT answers.
703,273 citations · 78,802 answers with sources · 21 brands · September 1 to October 18, 2026 · 7 platforms
What it means for you. If your buyers use Google's AI features, Gemini or Copilot, Reddit discussions about your category are part of what those answers read. Look at which threads are cited for your prompts and whether what they say about you is accurate. For ChatGPT and Claude, Reddit matters much less in our data.
What we don't know yet. The share depends on the industry. Consumer categories likely see more Reddit than business software, and our sample of 21 brands does not let us split by industry with confidence.
Last checked October 19, 2026.
Should brands post on Reddit to get into AI answers?
Take part where you would add something a buyer would want, and say who you are. Do not plant posts. The threads AI answers cite are the ones that already hold real discussion.
What we measured. See How much do AI answers cite Reddit? We measured where Reddit appears in AI answers. We have not measured the effect of a brand posting on Reddit.
What we found. Reddit appeared in 12.7% to 20.0% of answers on Google's AI features, Gemini, Copilot and Perplexity, and in 1.0% or less on ChatGPT and Claude (September 1 to October 18, 2026).
78,802 answers with sources · 21 brands · September 1 to October 18, 2026
What it means for you. Find the threads cited for your prompts. If they contain wrong facts about you, answer them openly as the company. If buyers ask questions your team can answer, answer them. Community rules and moderators remove promotional posts, and an answer that cites a removed or criticized post does you no good.
What we don't know yet. We have no measurement of whether posting raises mentions. A fair test would need many brands and months, and would have to avoid posts that break community rules.
Last checked October 19, 2026.
Do "best X" listicles get cited by AI?
Yes, in a minority of citations. About one in eight cited pages had "best," "top," "vs" or "alternatives" in its address. A closer reading of a smaller sample put listicles at about 6% of citations.
What we measured. We used two methods. First, we counted citations whose web address contained best, top, vs or alternatives, across 760,722 citations for 21 brands, September 1 to October 18, 2026. Second, Atlas classifies the most-cited pages for each brand by page type. That sample covers 1,285 pages and 28,629 citations across 11 brands.
What we found. 13.5% of citations went to addresses with those words. Microsoft Copilot was highest at 20.3%. ChatGPT was lowest at 4.0%. In the classified sample, listicle pages received 1,682 of 28,629 citations (5.9%). Review pages received 2,029 (7.1%) and informational pages 7,187 (25.1%).
760,722 citations · 21 brands · September 1 to October 18, 2026
What it means for you. Listicles are one source among many. Be accurate in the comparison pages that already rank for your category, and earn coverage from editors who review products in your category. We do not recommend paying to be inserted into lists. Platforms and readers both discount pages that exist to sell placements.
What we don't know yet. Address words are a rough signal. Many "best" pages are not listicles, and many listicles do not use those words. The classified sample covers each brand's most-cited pages only. We have not measured whether listicle citations are falling over time.
Last checked October 19, 2026.
Bots and access
Which AI bots visit a website, and what are they doing?
On our website, bots fetching pages for a person's live question were about as common as bots collecting training data. Bots building AI search indexes were the smallest group.
What we measured. Our CDN groups AI bot requests into three categories. Training crawlers collect pages to train models. Assistant fetchers retrieve a page because a person asked an AI assistant something. Search crawlers build the indexes AI search features use. The window is August 6 to October 19, 2026, on momenticmarketing.com.
What we found. The CDN counted 47,616 AI bot requests. Training crawlers made 24,888 (52.3%). Assistant fetchers made 19,072 (40.1%). Search crawlers made 3,656 (7.7%). The largest single bots were Meta's training crawler (14,653 requests), ChatGPT's user fetcher (13,325), Anthropic's ClaudeBot (4,074) and Amazonbot (3,674).
47,616 AI bot requests · Momentic's website · August 6 to October 19, 2026
What it means for you. A large share of AI bot traffic is retrieval for a person's question in that moment. Blocking or slowing those fetchers removes your pages from answers a person is waiting on. Look at your own bot logs by category before deciding what to block.
What we don't know yet. This is one website with a marketing audience. The mix will differ by industry and by how often a website changes. The CDN's category labels are its own classification.
Related research: AI user-agents, bots and crawlers to watch.
Last checked October 19, 2026.
Is GPTBot the crawler that matters for ChatGPT answers?
On our website, GPTBot was the least active of OpenAI's three bots. ChatGPT-User, which fetches pages while ChatGPT answers a person, made 35 times as many verified requests.
What we measured. OpenAI runs three bots. GPTBot collects training data. OAI-SearchBot builds the index ChatGPT search uses. ChatGPT-User fetches a page when ChatGPT needs it for a person's question. We counted requests our CDN verified as coming from each bot on momenticmarketing.com, August 31 to October 18, 2026.
What we found. ChatGPT-User made 6,155 verified requests. OAI-SearchBot made 2,401. GPTBot made 176. For Anthropic, ClaudeBot made 2,128 verified requests and Claude-User made 1,666.
Verified requests · Momentic's website · August 31 to October 18, 2026
What it means for you. Allow OAI-SearchBot and ChatGPT-User if you want ChatGPT to cite you, and test both in your robots.txt and firewall. A rule written only for GPTBot leaves the bots that serve live answers unaddressed.
What we don't know yet. This is one website. Verified counts exclude requests the CDN could not confirm. Unverified requests using these names were about as numerous, and some of those are impostors.
Related research: AI user-agents, bots and crawlers to watch.
Last checked October 19, 2026.
If robots.txt allows AI bots, can they reach my website?
Not always. Firewalls and bot protection can refuse AI bots that robots.txt allows. In our checks, refusals at that layer were more common than robots.txt blocks.
What we measured. For the same 155 websites as in Should I block AI training crawlers?, Atlas requested each homepage while presenting each of 38 AI bot identities. It recorded whether robots.txt allowed the bot and whether the server answered. The checks ran between September 25 and October 19, 2026.
What we found. Robots.txt blocked at least one bot on 9 websites. The server refused at least one bot that robots.txt allowed on 61 websites (39.4%). Across all website and bot pairs, 525 of 5,890 (8.9%) were refused at the server, and 131 (2.2%) were blocked by robots.txt. Refusals were more common on websites behind one large CDN: 41 of 86 (47.7%), against 20 of the other 69 (29.0%).
155 websites · 38 AI bots · checks run September 25 to October 19, 2026
What it means for you. Check your firewall and bot protection settings as well as robots.txt. A bot protection product can refuse AI bots even when your robots.txt welcomes them, and nobody on your team may know it is happening. Ask whoever runs your CDN which AI bots it challenges or blocks.
What we don't know yet. Our check comes from our own servers, not the AI companies' networks. Some firewalls refuse unknown IP addresses that present a bot's name, and let the real bot through. So a refusal in our check is a lead to investigate, not proof that the real bot is blocked. Your server logs settle it.
Related research: AI user-agents, bots and crawlers to watch.
Last checked October 19, 2026.
Should I block AI training crawlers?
It depends on what you sell. If your content is the product, blocking training crawlers can protect it. If you want AI platforms to know and describe your brand, we would allow them. Keep training and retrieval decisions separate either way.
What we measured. Atlas checks 38 AI bots against each monitored website's robots.txt. We looked at the latest check for 155 websites: brands Momentic measures and their competitors. The checks ran between September 25 and October 19, 2026.
What we found. Few of these websites block anything. Robots.txt blocked at least one AI bot on 9 of 155 websites (5.8%). Training crawlers were the most blocked group: 55 of 1,860 website and bot pairs (3.0%). Search crawlers were blocked in 33 of 1,550 pairs (2.1%). User-triggered assistant fetchers were blocked in 23 of 1,550 (1.5%).
155 websites · 38 AI bots · checks run September 25 to October 19, 2026
What it means for you. Most businesses in our sample let every AI bot in. Blocking a training crawler and blocking a retrieval bot have different costs. A training block affects what future models know. A retrieval block removes your pages from answers today. If you block, name each bot in robots.txt and check the result, because one broad rule can block both kinds.
What we don't know yet. We have not measured whether blocking training crawlers changes how AI answers describe a brand. That effect would take months to show up, because it depends on model release cycles.
Related research: AI user-agents, bots and crawlers to watch.
Last checked October 19, 2026.
llms.txt and files for agents
Do AI crawlers read llms.txt?
Rarely, on our website. AI crawlers fetched our robots.txt about 112 times as often as our llms.txt over 83 days.
What we measured. Momentic's website counts every request to /llms.txt and /robots.txt by the user agent the requester sends. The window is July 29 to October 19, 2026. We grouped requests that identify as OpenAI, Anthropic (Claude), Perplexity, Google, Microsoft or Meta.
What we found. Requests from those six companies fetched robots.txt 6,851 times and llms.txt 61 times. Claude's bots fetched robots.txt 2,495 times and llms.txt 6 times. OpenAI's bots fetched robots.txt 734 times and llms.txt 23 times. Perplexity's bots fetched robots.txt 463 times and llms.txt 5 times. A second count from our CDN's AI bot report agrees in direction. Between August 7 and October 19, it logged 3,752 AI bot requests for robots.txt and 6 for llms.txt.
Momentic's website · July 29 to October 19, 2026 · requests identifying as OpenAI, Anthropic, Perplexity, Google, Microsoft or Meta
What it means for you. An llms.txt file costs little to publish, and our data gives no reason to expect AI platforms to read it often. Put the work into the pages and files AI bots do fetch: robots.txt, your sitemap and the pages themselves. If you publish llms.txt, keep it accurate, because people and tools other than the large AI platforms read it.
What we don't know yet. This is one website. User agents are self-reported, and some bots do not identify themselves. We have not measured whether llms.txt changes how often AI answers mention or cite a brand. A study we have planned would measure that.
Last checked October 19, 2026.
Should I publish llms.txt, llms-full.txt or both?
Publish llms.txt if you want a short, accurate summary of your website for any tool that looks for one. We have no evidence that llms-full.txt adds anything, and we do not publish one.
What we measured. We counted requests to our own llms.txt for 83 days (see Do AI crawlers read llms.txt?). We do not publish llms-full.txt. A file that does not exist cannot be measured, and that is a gap in our evidence.
What we found. AI platforms' identified bots fetched our llms.txt 61 times in 83 days. They fetched robots.txt 6,851 times in the same period. People with browsers fetched llms.txt 633 times, more than all identified AI company bots together. llms.txt is a plain text summary of a website. llms-full.txt is a convention for putting a website's full text in one file.
Momentic's website · July 29 to October 19, 2026
What it means for you. A short llms.txt is cheap to keep current. A full-text file is a second copy of your website that you have to keep in sync, and stale copies give AI tools wrong facts. Until someone shows that AI platforms fetch and use llms-full.txt, we would put that effort into the pages themselves.
What we don't know yet. We have no data on llms-full.txt fetches or effects. A study we have planned would publish one on our website and a few volunteer websites and count who fetches it.
Last checked October 19, 2026.
Do AI agents ask for markdown versions of pages?
Some do, and it is a small share. Markdown was 3.6% of the responses our website served to AI bots.
What we measured. Momentic's website returns a markdown version of a page when a request asks for it. Our CDN's AI bot report counts responses to AI bots by content type. The window is August 7 to October 19, 2026.
What we found. Our website served 46,541 responses to AI bots. 39,075 were HTML (84.0%). 3,719 were plain text (8.0%), which includes robots.txt and llms.txt. 1,660 were markdown (3.6%). 1,371 were RSS feeds (2.9%).
46,541 responses to AI bots · Momentic's website · August 7 to October 19, 2026
What it means for you. HTML is still what AI bots fetch most from our website. Serving markdown on request is a low-cost option for the minority of agents that ask, and it does not replace well-structured HTML.
What we don't know yet. The content-type report has no bot breakdown, so we cannot say which agents asked for markdown. We have not tested whether serving markdown changes citations. This is one website.
Last checked October 19, 2026.
Do agents use MCP server cards, agent skills indexes and API catalogs?
Our MCP server gets steady use. The discovery files that point agents to it get very little.
What we measured. Momentic's website counts requests to each agent-facing file by user agent. The window is July 29 to October 19, 2026 (83 days). MCP is a standard way for AI tools to call a website's functions. The server card, skills index and API catalog are files that tell agents those functions exist.
What we found. The MCP endpoint received 79,008 requests. Most came from software clients that do not name an AI company. Requests identifying as Claude made 739 of them. The discovery files received far fewer requests: the MCP server card 285, the API catalog 217, agents.md 190, auth.md 190 and the agent skills index 161. Requests identifying as OpenAI, Anthropic, Perplexity, Google, Microsoft or Meta fetched the server card 40 times in 83 days.
79,008 MCP requests · Momentic's website · July 29 to October 19, 2026
What it means for you. If you have functions an agent can call, an MCP server can get real use. In our data, agents reached it without reading the discovery files first, likely through directories and direct configuration. Publish the discovery files, since they are small, and do not expect them to bring traffic on their own.
What we don't know yet. We cannot see how clients found our MCP server. Many requests may come from the same few clients. This is one website with an MCP server built for marketers.
Last checked October 19, 2026.
Content and pages
Does updating content or dates raise AI citations?
We do not know yet from our own data. The signals we have point toward keeping pages current, and we would not change dates without changing the content.
What we measured. We measured how often AI platforms add a year to their searches (see Do AI platforms add the year to their searches?). We have not measured the age of cited pages or run a test of updating pages.
What we found. Claude added 2025 or 2026 to 33.7% of its searches, ChatGPT to 8.5% and Gemini to 1.3%, from September 1 to October 18, 2026.
96,555 searches · 21 brands · September 1 to October 18, 2026
What it means for you. Update pages when the facts change, show the date of the last real update, and keep visible dates consistent with the dates in your page code. Changing a date without changing the content gives platforms and readers a false signal.
What we don't know yet. Whether recent pages get cited more, and how fast that falls off. A study we have planned would measure the age of pages AI platforms cite against the pages ranking for the same prompts.
Last checked October 19, 2026.
Do schema, author bylines or longer content get a page cited by AI?
In our data, not in a way we could detect. Among pages that already rank in Google's top 5 for a prompt, the pages AI platforms cited and the pages they skipped had the same schema markup, author signals, length and headings. Pages with dates were cited somewhat more often, and the cited rate differed widely by type of website.
What we measured. We compared 139 pairs of pages for 21 brands, from September and October 2026. Each pair was two pages from Google's top 5 for the same prompt, one position apart. AI platforms cited one page and not the other. In half the pairs the cited page ranked higher, and in half it ranked lower. We fetched both pages and recorded schema markup, dates, word count, author signals, headings and tables. We also looked at all 55,830 top 5 results in the same period to compare types of website.
What we found. Schema markup was on 81.3% of cited pages and 77.7% of uncited pages. An author signal was on 65.5% and 66.2%. Median length was 1,707 words for cited pages and 1,811 for uncited pages. None of these gaps was larger than chance. Dates showed a smaller gap that may also be chance: 61.2% of cited pages had a date in their markup, against 51.1% of uncited pages. When both pages showed a date, the cited page was the newer one 63% of the time. The type of website showed a large gap. Top 5 results on Facebook, LinkedIn or Instagram were cited by an AI platform 20.1% of the time. Results on other websites were cited 46.4% of the time, at every ranking position and for 20 of 21 brands. On ChatGPT, Perplexity, Claude and Copilot, those social pages were cited 0.2% to 2.4% of the time.
139 page pairs and 55,830 top 5 results · 21 brands · September and October 2026
What it means for you. For a page that already ranks, adding schema, a byline or more words is not likely to be the change that gets it cited. Keep those for their other benefits. Show a clear published or updated date (see Does updating content or dates raise AI citations?). Put the content you want AI platforms to cite on your own website, and use social posts to point to it. Facebook and Instagram block the ChatGPT, Perplexity and Claude search crawlers in their robots.txt files. Copilot also cited these pages rarely although Bing's crawler is allowed, so robots.txt is one reason among others, such as login walls.
What we don't know yet. Our sample can detect gaps of about 10 to 15 points, not smaller ones. We compared pages that already rank in Google's top 5, so this does not cover pages outside it. We measured pages after they were cited, and some may have changed since. These are associations. We have not tested whether changing a page changes its citations.
Related research: Re: GEO, AIO, AEO, SEO.
Last checked October 19, 2026.
Questions we are still studying
We do not have our own data on these yet. Each one has a study planned.
- Does an llms.txt file raise how often AI mentions or cites a brand? We plan to publish llms.txt on half of a matched set of websites and compare mention and citation rates before and after.
- Is pay-per-crawl worth turning on? We plan to turn on pay-per-crawl for one test section of our website and track AI bot requests, status codes and citations.
- Do AI crawlers run JavaScript? We plan to publish test pages whose key facts appear only after JavaScript runs, and check which bots fetch the scripts and which answers quote the facts.
- How much of my traffic comes from AI? We plan to compare AI referral sessions with organic, Direct and total sessions for the brands with Google Analytics connected.
- Should tracked prompts read like personas or like keywords? We plan to run matched pairs of persona-style and keyword-style prompts with the same intent and compare the answers.
- Do AI Overviews reduce clicks from Google? We plan to compare Search Console click-through rates at matched positions with and without an AI Overview.
- Does IndexNow make AI platforms cite new pages sooner? We plan to compare the days from publishing to first AI citation for pages submitted through IndexNow and pages that were not.
- Does publishing more pages raise how often AI recommends a brand? We plan to compare mention and recommendation rates before and after brands publish new pages, against matched brands that did not.
- Can I tell when an agentic browser visits my website? We plan to analyze our own website's logs for agent-driven sessions and run test visits from known agent browsers.
