Prompt tracking means asking ChatGPT, Gemini, Perplexity, and Claude the same questions on a schedule and recording whether your brand shows up, whether your pages get cited, and whether you get recommended. Most advice about it is untested: write longer prompts, aim for a certain character count, test each one a few times by hand. We wanted to know what actually changes the answers.
So we measured it. Between July 13 and August 15, 2026, our tracking ran 8,342 prompts for 18 brands across 11 answer surfaces (ChatGPT, Claude, Gemini, Perplexity, Google AI Overviews, Google AI Mode, Bing Copilot, Google Search, and the no-search API modes of the chat assistants), producing 156,233 answers, 145,498 of which were scorable. Every scored answer records whether the brand was named, which other options were named, what got cited, the sentiment, and any caveats the assistant attached. The numbers below come from that table. The method at the end is built on them.
Brand questions and category questions are two different instruments
Across the corpus, prompts that name the brand ("Would you recommend [brand]?", "Is [brand] legit?") got the brand into the answer 91.6% of the time (57,824 answers). Prompts that don't name it got the brand in 19.6% of the time (87,674 answers). Averaging those into one number produces a statistic that describes nothing. Give brand prompts their own group and keep them out of the number you report upward.
But don't treat brand prompts as throwaway inflation, because the interesting part is the 8.4% where the brand asked about by name still doesn't come back cleanly. Per brand, the floor varies more than you'd guess: the median brand in our corpus was named in 95.7% of answers about itself, and the lowest was named in 66.5% (16 brands with at least 200 scored answers each).
The brand at 66.5% is worth describing, anonymized, because the failure mode is instructive. It's a services firm whose name is also an ordinary industry phrase, the equivalent of an agency called Precision Marketing. Ask "Are there any downsides or risks to using [the brand]?" and the assistants answer about the technique instead of the company: a tidy rundown of the risks of the marketing approach that happens to share the brand's name, with the company nowhere in it. Its comparison prompts fail the same way; one assistant opened by explaining that the question was a category error, since "you don't choose between a strategy and an agency." A third of everything the machines say when asked about this brand is about a concept, and no amount of publishing fixes that until the naming and entity problem gets fixed first. A brand prompt is a smoke detector. It's boring right up until the day it isn't, and when it goes off, that finding outranks everything else in the report.
The ask changes the answer more than the length does
There's a popular claim that tracked prompts should land between 180 and 210 characters. Our corpus holds few prompts above 200 characters, so we can't test that exact range. We can test what the length is supposed to buy: answers that name options and make a pick. That turns out to come from the ask, and the effect is large.
Non-branded prompts that explicitly ask for a pick ("what would you recommend," "best X for Y," "which should I choose") produced answers naming 4.55 options on average, against 1.25 for prompts that ask about the topic without asking for a pick (24,225 vs 63,455 answers). The tracked brand landed a recommendation in 22.8% of pick-ask answers, against 8.0% otherwise, with brands and platforms held constant. If your prompt doesn't ask for a recommendation, most of what comes back is explanation, and you can't measure whether you're the choice in an answer that never chooses.
The cleanest illustration comes from tracking our own agency. "Best alternatives to [a well-known competitor]" produced answers naming 12.5 options on average across 51 runs. "What KPIs should an SEO agency report on each month?", a real buyer question we also track, named zero options across 50 runs. Both rows are useful. The first can measure whether we get picked; the second can only ever measure whether our thinking shows up in the explanation, and mixing the two into one visibility number hides which one moved.
Length is a proxy for something real, which is why the folk advice half-works. A prompt needs four parts: who is asking, their situation, the specific problem, and the ask. Carrying those usually takes a sentence or two. Count the parts and let the character count land wherever it lands.
Short prompts have a separate job, and the data backs keeping a few. Category-head prompts under 40 characters ("best [category]") elicited list-style answers naming 2.68 options on average, and mentioned the tracked brand more often than mid-length prompts did (30.4% vs 17.3%). They don't simulate a buyer, and they aren't supposed to. They test whether the machines already treat you as a default in your category. A handful of them is a useful dashboard row; a set built out of them is a keyword list wearing a question mark.
One objection to context-rich prompts deserves a straight answer: nobody types a 200-character message into ChatGPT. True. But a real buyer gets there across several messages, describing their situation and following up before asking what to pick. By the time the recommendation question lands, the model has the context from earlier in the conversation. Tracking tools send one message, so your prompt has to compress that accumulated context into the final ask. You're simulating the end of the conversation, because the end is where the decision gets made.
One in twenty identical prompts changes its mind the same day
We run many prompts multiple times per platform within the same run. Across 29,344 of these repeat groups (same prompt, same platform, same day), the answers disagreed about whether the brand was mentioned at all 4.5% of the time on non-branded prompts and 5.1% on branded ones.
That number should change how you read tracking data. A concrete example from the corpus, anonymized: "What is the best car wash in [a mid-size Western city]?" ran three times on Perplexity inside one run. The brand we track was absent from the first answer and present in the second and third. Same question, same platform, same afternoon. A brand "appearing" this week and "disappearing" next week is, one time in twenty, the same answer wobbling. This is why testing a prompt once before tracking it proves close to nothing, and why any tool reporting week-over-week movement on single runs is selling you noise.
It goes deeper than mentions. In a separate experiment we ran a single prompt ten times through Gemini with search grounding and captured the searches the model itself issued behind each answer. No individual search string appeared in all ten runs, and 86.8% appeared only once. But the themes those searches covered were identical in every run: same entities, same attributes, different words each time. The machine is consistent about what it checks and inconsistent about how it phrases the checking. So judge prompts, and read reports, at the theme level. String-level data shuffled in every experiment we've run; the theme level held still in every one.
Practically: test each candidate prompt with repeats, keep the ones that produce a measurable, sourced answer consistently, and distrust any conclusion drawn from one run of anything.
The question writes the answer
This finding should worry anyone comparing brands on sentiment. We split our branded prompts by framing: prompts with negative pull ("Is [brand] a scam?", "Are there hidden fees?", "What problems do people have with [brand]?") against everything else.
On negative-framed prompts, 67.1% of answers attached a caveat to the brand, and average sentiment scored slightly negative (7,859 answers). On the rest, 32.0% attached a caveat and average sentiment scored clearly positive (49,965 answers). Same brands, same platforms, same period. The framing of the question roughly doubled the caveat rate and flipped the sentiment sign.
Neither number is wrong. Ask about problems and the assistant dutifully finds problems. Row by row, anonymized, it looks like this. "Are there any hidden fees with [brand]?" pointed at a live-events venue returns an itemized list of every surcharge the assistant can find. "Is [brand] worth the price?" pointed at a specialty lighting manufacturer produced an answer that praised the products and still attached a caveat that pricing can run higher than mass-market alternatives; the question demands a yes-but, so the answer manufactures one. "Did [brand] raise its prices recently?" drew, from an assistant with no browsing and no data either way, a non-answer framed entirely around pricing uncertainty, and that scores as a caveat too. Sometimes the caveat is even true; the venue's fees are real. The measurement problem is the aggregate: a set loaded with these questions shows roughly double the caveat rate of a balanced set no matter which brand it points at. The error happens at the reporting layer, when a set that skews negative gets compared against a set that doesn't, and the difference gets presented as a difference between brands. It's a difference between questionnaires.
Two rules follow. First, classify every prompt's framing as negative, neutral, or positive before it enters tracking, and balance the set to roughly net zero, especially in any set used to compare brands against each other. Second, check that every question makes sense for the specific business it measures. A hidden-fees question is a real question for a subscription business and a manufactured negative for a business that sells one thing at a posted price. We've watched a template set manufacture a false winner in a benchmark this way. A question that doesn't fit the business produces an artifact, and artifacts get presented as findings.
Platforms split into citation regimes, and it decides what you can influence
Whether an answer cites sources determines whether your publishing can move it. Our corpus splits cleanly:
- Bing Copilot and Google Search: sources on ~100% of answers
- Perplexity, Gemini with search, and Claude with search: ~99%
- Google AI Mode: 96%
- Google AI Overviews: 79%
- ChatGPT with browsing enabled: 88%, once we corrected for our own bug. The whole-corpus number read 23%, and pulling raw payloads showed why: a defect in our pipeline truncated ChatGPT's answers before the citations attached. After the August 13 fix, 508 of 575 answers carried sources
- API-mode ChatGPT, Claude, and Gemini without search: 0%, by design
Two lessons for prompt testing. First, "did it search" tells you little, because searching and citing are separate behaviors; read what actually got cited. If a prompt draws citations from comparison pages, community threads, and editorial writeups, those are surfaces you can plausibly appear on, and the prompt is worth tracking. If it cites nothing, or only encyclopedia-style pages, you have no citation to win and no way to verify what shaped the answer. Second, a rate that sits far outside its platform's band deserves an audit of your own pipeline before it becomes a finding. Ours did, and the outlier was us.
The method, given all that
Where the wording comes from. Your own content is the worst source; it's written in your internal vocabulary, and buyers don't use it. Pull phrasing from sales calls, onboarding calls, and support tickets, where prospects describe problems in their own words. Pull long queries from Search Console with the regex filter ([^" ]*\s){5,} (six or more words) over a 6 to 12 month window, and note where searchers' vocabulary differs from your website's; you need prompts in the searchers' vocabulary, and probably content too. If your tooling captures them, mine the searches the assistants themselves run behind answers in your category, at the theme level, since the strings shuffle. Community threads round it out. Then compress each raw phrasing into the four parts: who, situation, problem, ask.
Testing before tracking. Run each candidate several times per platform you'll track. Keep it if it consistently produces a relevant answer, draws citations from surfaces you could appear on, and names options when it's supposed to. Rework or cut it if it comes back generic, sourceless, or unstable. Our flip-rate data says one clean answer proves little; consistency across repeats is the bar.
Balance and sense. Classify framing per prompt, net the set to roughly zero, and confirm every question is answerable for this specific business. Ten minutes of this prevents the instrument effects above from becoming your headline findings.
Mix and size. Start around 30 to 50 prompts: mostly full-context prompts with an ask, a few short category checks, and brand prompts in their own group. Then let your demand data reshape the mix; if Search Console shows question-style demand piling up in one product area, the set should lean there. And pair each prompt with the page on your website that should win it. A prompt with no plausible page is a content decision first and a tracking row second. The pairing is what turns a report into a work queue.
Cadence. Baseline for one to two weeks before the first report. Review at two to four weeks with someone who knows the customers, because some prompts that tested clean will fail under real tracking; our own kick-off reviews cut or rework prompts every time. Refresh quarterly from new call transcripts, new long queries, and newly captured machine-issued searches. A prompt set that hasn't changed in six months is measuring last year's conversation.
The checklist
- Wording sourced from calls, tickets, and Search Console long queries, in buyer vocabulary
- Every prompt carries who, situation, problem, and an ask; prompts meant to measure recommendation actually ask for a pick
- Every prompt makes sense for the specific business it measures
- Framing classified per prompt; the set nets to roughly zero
- Every prompt tested with repeats per platform; judged on relevance, citations, named options, and consistency
- Brand prompts grouped separately, read as a smoke detector, excluded from the headline number
- Every prompt paired with the page that should win it, or flagged as a content gap
- Baseline window set; two-to-four-week review scheduled; quarterly refresh on the calendar
The tool you track with matters less than what you feed it. Feed it questions a buyer would ask, balanced so the questionnaire isn't the finding, tested until they're stable, and tied to pages you can actually change.
this text
