Momentic engineered Atlas to measure AI answers and act on them.
Atlas is an intelligence and testing system the Momentic team uses to understand and influence what AI platforms say about a brand, its products and services, and its category. Atlas is best described as collective team intelligence: the measurements it takes, the judgments Momentic strategists record inside it, and the rules the team writes from those judgments accumulate in one system. It learns and adjusts for as long as the team keeps working inside it, and this page explains how.
What is Atlas by Momentic?
Atlas is a strategic intelligence and testing tool the Momentic team operates for clients who want to know, specifically, what AI platforms say about them and why. Momentic engineered Atlas on retrieval science, testing, and statistics. It analyzes content passage signals, real query fan-outs, embeddings, mentions, citations, recommendations, sentiment, prominence, content updates, and entities, across answers that cite news sites, forums like Reddit, review platforms like Yelp, and social platforms. Atlas has catalogued more than 500,000 unique sources cited in AI answers. Atlas is not a dashboard.
Who is Atlas for?
Atlas is operated for brands that want to know what AI platforms say about them, their products, and their category, and want to change it. Today that is Momentic's clients across healthcare, B2C, ecommerce, and SaaS. Inside those companies, four groups read it: marketing teams working the opportunity backlog, growth teams working the demand data, product teams reading what platforms warn buyers about, and leadership reading the competitive picture. The team section below describes what each group gets.
Why did we engineer Atlas?
Our clients demand AI intelligence now, for decisions they are making now. We built Atlas so our clients would not have to wait for generally available tools to be accessible. In addition to this, it is our firm belief that a self-service SaaS tool adds too many restrictions for proper customization of data, insights, analysis, thus actions.
The first 30 days of Atlas
Over the past 30 days, Atlas processed and analyzed more than 1 million AI citations and more than 25,000 AI recommendations for healthcare, B2C, ecommerce, and SaaS clients. A citation is a source an AI answer used. A recommendation is an answer presenting a brand as an option to choose. Atlas builds this record from grounding queries, real query fan-outs, and first-party analytics. Query fan-outs are the follow-up searches a platform runs on its own while composing an answer, and Atlas captures them from live answers.
Since July 1, Atlas has scored answers from Google's AI Overviews and AI Mode, classic search results, ChatGPT, Claude, Gemini, and Perplexity. We've meticulously tested with no web-search tools and with web-search tools.
Answer stability: AI platforms are not consistent with their answers
No matter the configuration, AI platforms are not consistent in their answers. The dimension to measure is answer stability, and the published numbers depend on how much of the answer you compare. SparkToro's study of 2,961 prompts across ChatGPT, Claude, and Google AI Overviews found the odds of getting the same brand-recommendation list twice were under 1 in 100, and the same list in the same order closer to 1 in 1,000. Our corpus measures a narrower unit: whether one brand's mention stays the same when the same prompt runs on the same platform on the same day. Across 29,344 of those repeat groups, about 1 in 20 contained both an answer that named the brand and an answer that left it out. Stated as stability, about 95% of repeat groups agreed on whether the brand was mentioned.
The two studies describe the same behavior at different units. The more of an answer you compare, the less stable it is: a full ranked list almost never repeats, while a single brand's mention stays the same in about 19 of 20 repeat groups. The same SparkToro data shows appearance rates vary far less than single answers: their top brands appeared in 55 to 77% of responses regardless of phrasing. Academic work on measuring AI visibility reached the same instruction and put it in the title: don't measure once. That is the design conclusion Atlas is built on. A brand's AI visibility is a distribution, not a fact. Atlas re-runs prompts, reports the rate at which answers agree, and treats a single answer as an unstable reading. We'd expect some inconsistency, since the platforms sample their answers rather than computing one fixed answer per prompt.
Splitting the same measurement by platform shows stability rises with web grounding. Across same-day repeat groups per platform (between 816 and 2,926 groups each), Google's search-grounded surfaces agreed on the brand's mention in about 99% of groups: AI Mode 99.6%, classic search results 99.1%, Copilot 99.1%, and AI Overviews 98.9%. Chat models answering with web access agreed in 92.7% to 95.9% of groups. The same models answering without web access agreed least, at 88.1% to 88.7%. At the stricter unit of the full set of brands an answer named, identical sets within a group ranged from 59.8% to 91.2% by platform. Stability is platform-specific, so Atlas reports it per platform rather than as one number. I'll keep publishing our findings as the corpus grows.
Atlas data contributed to decisions in board rooms and on growth, product, and marketing teams in its first month. The next section describes what each team gets.
How Atlas learns from the team that runs it
Prompt tracking is one function of Atlas. The function that gains value over time is the loop between machine measurement and recorded human judgment. Atlas measures answers and proposes: candidate prompts with the evidence that justifies them, opportunities with scores, content drafts, and competitors that began appearing in answers. A person on the Momentic team approves, reworks, or dismisses each one, and Atlas records the decision with its reason. The team writes rules from the patterns in those decisions, and each rule then runs automatically on every brand.
One example of that loop closing. When we caught a shared template asking restaurants about subscription pricing, deleting those prompts would have fixed one brand once. Instead the fix became a business-fit check that now validates every prompt against the brand's declared business model before it runs. A mistake the team catches once becomes a check Atlas runs from then on.
The team's knowledge is stored in Atlas as data its algorithms read. Buyer personas from sales calls, a brand's business model, its competitor list, the rules for separating a brand from a similarly named company, customer language pasted from calls and support tickets: team members record these while doing client work, and Atlas's generation and scoring algorithms consume them directly. What a strategist knows becomes configuration the system executes.
Atlas also uses its own output as input. The follow-up searches platforms run while answering tracked prompts become evidence for the next quarter's candidate prompts. The pages platforms cite become outreach targets. The warnings platforms attach to a brand become content briefs. The scheduled reviews retire prompts that no longer produce scoreable answers and draft replacements from what changed.
This loop is why Atlas runs as a managed system. As the team works inside it, approving verdicts, entering customer language, dismissing weak opportunities, and correcting classifications, the record each brand is measured against grows. Atlas holds what the team has learned in a form that runs automatically, and its rules and evidence grow for as long as the team keeps using it.
How teams use Atlas
Marketing teams work from the opportunity backlog. Atlas turns its measurements into a prioritized list of actions, each backed by the evidence that produced it, in three categories: create a page for a prompt no brand currently answers, refresh a page that ranks just below page one or gets cited without being recommended, and pursue a mention from a specific third-party website that platforms cite on questions where the brand is absent. For Momentic's own brand, the backlog holds 385 open items.
The citations view reports the exact domains and pages that AI platforms cited for each prompt, and it classifies each one as the brand's own, a competitor's, or third party, so outreach targets a specific page instead of "earn more mentions."
A content refresh workflow turns a flagged page into a brief draft, and a person approves or rejects the draft. Upon approval, it goes to a writer.
Growth teams use the demand side of Atlas. Atlas ranks each tracked prompt by demand, computed from the brand's own Search Console and keyword data, so a coverage gap on a high-demand prompt outranks one on a low-demand prompt. It reports AI referral traffic from GA4 alongside those gaps. And for brands on Cloudflare, Atlas ingests bot analytics to show which pages AI crawlers fetch, which pages they skip, and whether the fetched pages are the ones answers actually cite.
Product teams read the caveat clusters in Atlas. When a platform hedges about a brand, Atlas records the warning, clusters the repeated ones, and keeps a verbatim example of each, split by whether the prompt named the brand. The clusters read like unsolicited field research: what AI tells buyers to check before choosing you, sourced from what it read about you. Teams compare these against their own survey findings and roadmap.
Leadership gets the competitive picture with denominators attached. For Momentic's own brand, 912 of 1,008 answers to comparison prompts mentioned at least one other company, and Atlas counts which companies, how often, and on which topics. When a company we do not track starts appearing in those answers, Atlas flags it as a competitor worth adding. Brand-prompt results are reported separately as a health floor, and a platform omitting the brand from an answer about the brand is reported as an emergency finding, not averaged away.
Analysts and other AI tools connect to the same data over MCP, an open protocol for tool access. The Momentic team runs its own agents against Atlas this way, which is what "plugging Atlas into other AI workflows" means in practice: the same queries, tools, and records, available to whatever system asks.
It balances the sentiment of each prompt set before comparing brands, because skewed phrasing lowers a brand's measured sentiment on its own. And its algorithms write prompts from recorded evidence: Search Console queries, the platforms' own fan-out searches, and the language customers use on calls, in tickets, and in reviews.
Atlas also does not claim what it cannot show. It does not tie a citation to revenue, because we have not seen a defensible way to measure that.
Our clients do not want scaled AI slop content. They want more informed recommendations, quicker, grounded in their category, their goals, and their brand.
What's next?
We will publish more on the methodology, the technology, and the findings.
How do I get access to Atlas by Momentic?
Currently, Atlas is available to anyone and everyone as a managed service or as a custom implementation of the Atlas Core into your systems. Self service SaaS isn't a feasible path for an intelligence system that needs many customizations based on nearly infinite variables. And a version of Atlas without the team using it every day would stop learning, because the loop that improves it runs on the team's recorded judgment.
For fun: Different Atlas logos we've been playing around with
