← All posts

A Brand Mention in AI Search Is a Count, Not a Rank

← All posts

One query, three passes on ChatGPT, three different lists of domains, and one domain that was in all three. The whole shape of the problem is in that sentence. Tracking brand mentions in AI search is a fixed set of buyer queries run on several engines, each answer split into domains the engine cited, domains it only named, and domains that were absent, then repeated, so you know what comes back a second time. There is no position in it. Across 2,961 runs of twelve recommendation prompts, ChatGPT and Google's AI answers returned the same brand list under 1% of the time and the same list in the same order under 0.1% (SparkToro and Gumshoe, January 2026). A single pass still answers a question, who is in the answer today, and for a first look it is often enough. What one pass cannot tell you is whether that domain is there every time, and the methods below are mostly ways of finding that out without pretending the number is more precise than it is.

A brand mention in AI search is a count, not a rank

A brand mention in AI search is an answer that names the brand, with or without a link. A citation is an answer that links to the brand's domain as a source. Absent is the third state, and on most queries it is the common one. Those three, counted per engine and per pass, are what tracking produces; the average position some trackers report is an average of a value that does not repeat.

Mike Sonders ran twelve prompts a hundred times each through logged-out ChatGPT. Each answer named about ten brands; across a hundred answers the pool was about 44, and only around five brands per prompt appeared in 80% or more of them. The others rotate through. In the SparkToro data, City of Hope appeared in 69 of 71 ChatGPT answers about West Coast cancer care and came first in only 25 of them. Presence is the stable thing. Order is close to noise, which is why "we moved from third to second" is not a result anyone can act on.

The Rank Tracker checks Google positions when you run it, and a position there is a thing with a history. An AI answer has a consideration set: a few brands the engine keeps returning and a longer tail it draws from when the context shifts. What a count can establish is which of the two your domain is in. Why it landed in one rather than the other is a different question, and the report does not answer it on its own.

Why mentioned and cited are two columns, not one score

Claude, in our reports, never returns a link. We query it through the API without web search, so what it names is what it remembers from training. Google AI Mode sits at the other end: it pulls pages from Google's index for nearly every answer and links them, so a mention there almost always comes with a URL. The two counts describe different things, and merging them into one visibility score removes the only part of the number you can act on.

EngineWhere the answer comes fromLinks in the answerWhat a mention there means
ChatGPTTraining data; the live web when it searchesOnly when it searches; clicks carry utm_source=chatgpt.comMemory, unless a source link is attached
Claude (via API)Training data onlyNoneThe brand is a known entity; no page involved
GeminiGoogle's own retrievalVaries by query; times out on someMixed; read the cited URL when there is one
Google AI ModeGoogle's indexSource links on nearly every answerA page was pulled for the answer

One vendor audit of 750 answers in March 2026 put the share of answers with at least one source link at 71% for AI Overviews, 62% for Perplexity, 41% for ChatGPT with search on, 28% for Gemini and 24% for Copilot (Vismore, which sells a tracker; its raw data is available on request). The exact figures will have moved since; the spread is the point. A brand that is cited on Google AI Mode and only named on ChatGPT has two different problems, one about a page and one about the entity, and a report that shows 1.5 mentions per answer across both engines has hidden which.

Three failure modes belong here and get one line each. A cited URL that does not contain the claim, or does not exist, counts as a citation in most tools and should not. A mention with last year's pricing is a memory answer about a stale entity, and no new page changes it quickly. And a domain named without a URL in a Google AI Mode answer, where URLs are the norm, usually means the engine knows the brand and found no page worth citing for that query.

How many passes: what one pass shows, what three separate, what none of them prove

One pass is often enough. We take that position on the tool, against the "run it ten times" guides. The difficulty with AI answers is retrieval: it is contextual, so the second pass can pull a different set of pages and a different set of domains, and both passes were right about what the engine did that time. Repeating the query does not make the number more accurate. It separates the domains the engine returns regardless of context from the domains that depend on it.

What a stability score is

Stability counts the passes, out of 1, 3 or 5, in which a domain came back for the same query on the same engine. A 3/3 next to a domain means the engine recommends it from memory or finds it every time; 1/3 means it turned up once and may not again. Each pass costs 5 credits per query per engine, so five queries on three engines at 3x is 225 credits. A repeat of the same run inside 24 hours is answered from cache and costs less; skipping the cache buys fresh answers at the full price. Which answers go in the denominator matters too. A pass that timed out is left out of the count. An answer that names no brands at all counts as a pass in which no one, you included, appeared.

What the research says about repeats

Three studies from this year put numbers on it, and none of them flatter three passes. Dmitrij Żatuchin's reanalysis of five brand-recommendation audits (arXiv 2609.04047, September 2026) puts the reliability of a single prompt's result at 0.58 after five iterations, 0.74 after ten and 0.81 after fifteen, and calls those tiers exploratory, confirmatory and rigorous. Sonders, after his 1,200 runs, suggested running key prompts about five times whenever you pull data, and a hundred or more for a rate you can call statistically meaningful. The University of St. Gallen team behind "Don't Measure Once" (Schulte, Bleeker and Kaufmann, arXiv 2604.07585, April 2026) puts the bar for claiming a brand has no visibility at around sixty runs: zero mentions in twenty runs is still consistent with a true rate near 16%.

PassesSingle-prompt reliability (Żatuchin, 2026)What it supports
1not measurableWho was in the answer today
3below 0.58Fixed domain vs rotating domain for this query
50.58A mention rate across a portfolio of queries, not per query
10–150.74–0.81A per-query rate you could put in a client report
60+not covered (St. Gallen)A claim that the brand is absent

Three passes do not make a single query's number reliable; the measured coefficient at five is 0.58. They tell you whether a domain is fixed or rotating. Declaring a brand absent takes sixty runs or more.

So we do not claim that 3x is more accurate than 1x. The stability counter is the raw count of passes the domain survived, shown next to the same count for every competitor in the answer. On five queries across four engines that is twenty cells. A 2/3 in one of them against a 3/3 in the next is the comparison that holds up; a mention rate computed from those twenty cells is one the research above would not sign off on. The spread between passes on one engine is also, as the next section shows, about the size of the spread between engines.

Engines disagree, so the report is per engine

BrightEdge compared the top-30 brands named by ChatGPT, Perplexity, Gemini, Google AI Mode and AI Overviews, pairwise, across industries (May 2026). Overlap between any two engines fell in a 36% to 55% band. In retail it was 97% and in travel 94%; in finance 71%, and in healthcare 60%, with some pairs as low as 40%. Cleveland Clinic took 6.5% of healthcare brand mentions on Perplexity and 0.3% on ChatGPT, a 22-fold gap for the same brand on the same queries. The St. Gallen data gives the same-day overlap of brand sets as 0.49 for Perplexity, 0.44 for ChatGPT, 0.41 for Gemini and 0.38 for AI Mode. An engine agrees with itself from one pass to the next about as often as two engines agree with each other in a research-heavy category.

A share of voice is therefore read per engine. Ours is counted against the competitors that appeared in the same answers, and each appearance is weighted lightly by where it sits in the answer, from 1.0 for the first brand down to a floor of 0.3, so a brand named last still counts. That weight is the only use the report makes of order, and it never shows a rank. The report gives share of voice per engine and for the whole run; the whole-run figure is a summary, and the per-engine one is where a decision comes from. "We are at 30% in ChatGPT" says nothing about Gemini. The transactional categories are the exception: when overlap is 97%, as in retail, the engines are mostly reading the same review sites, and the per-engine view will look redundant. It stops looking redundant the moment one engine switches from memory to search.

The line in a report that shows this most often is Gemini citing two competitors while ChatGPT and AI Mode cite you. Nothing there needs resolving: two retrieval paths landed on different pages, and the fix, if there is one, is specific to the engine that did not find yours.

Building the query set: five lines, fan-out, and names that fool a count

The queries are the sentences a buyer types: best X for Y, X vs Y, is X worth it for a team of five, what does X cost. A run takes up to five of them, one per line, and the fan-out option expands a topic into the sub-queries an engine would split it into, up to fifteen. Sales calls, the People also ask box and the Reddit threads in the category are where the phrasing comes from; the point is to track what gets asked.

Prompt volume is the number everyone wants next, and nobody has it. OpenAI publishes no per-query counts, and the prompt volumes some suites report are modelled estimates, not counts from the engines. The report here does not show a volume per query because there is no measured one to show. Rank the five lines by what they mean to the business, and accept that how often people ask them is unknown.

Google AI Mode, one of the four engines here, answers commercial queries as readily as informational ones and links its sources on nearly every answer, so a "best X for Y" line is not wasted on it. AI Overviews, the summary block on the ordinary results page, are a different product: whether one fires on your queries is a question for a rank check, not for this report. One more thing to check before reading any count: a brand that shares its name with a common word, a place or another product will be over-counted by any tool, ours included, until the answer text is read.

Reading the report: three counters and what to do with each

I read the two counters at the top before anything else, because together they say what kind of query this is. In the sample report of our LLM Rank Tracker, 72% of references are citations and 28% are mentions without a link. That is a search-driven query: the engines went to the web, a page took each slot, and a page can take it back. Expect rotation between passes on these queries, since retrieval is where the context changes. When the proportions flip and mentions outnumber citations, the engines answered from memory. The answer will be stable from pass to pass, and no page you publish next month will move it quickly; what moves it is the brand's record as an entity across the sites the models were trained on, which is a different piece of work.

What the report showsWhat it meansWhere to look next
Cited 72%, mentioned 28%Search-driven query; the slot belongs to a pageThe competitor's cited URL and title in the table
Mentioned above citedMemory answer; stable, and slow to changeEntity consistency, not a new page
Domain types: 46% review sites, 31% vendors, 23% blogsShortlists hold the slotA line in those shortlists; your own guide will not take it
Your domain 2/3, a competitor 3/3You are in rotation; they are fixedWhat their URL has that yours does not
Gemini cites them, ChatGPT and AI Mode cite youEngines disagree; nothing is brokenGemini's cited URL only

The domain-type split is the one most people skip, and it decides the strategy. When half the references are review sites, the engines are reading a shortlist someone else wrote, and a guide on your own domain is competing with the wrong page type. When vendors dominate, the engines are comparing product pages directly, and the cited URL column shows which page per competitor. The 2/3 against 3/3 comparison belongs to the same table: the domain that came back three times has a page the engine returns to regardless of context, and its title is right there. What the report cannot show is why that page and not yours. That takes reading both.

The report shows every answer in full with its numbered references, so the count can be checked by hand, and the CSV and Excel exports carry each domain with its URL, title, citation flag and stability count.

Manual spreadsheet, API build, or a tracker: what each one counts

All three approaches run the same loop: a fixed query list, several engines, a log of who was named, who was linked, and under what conditions. They differ in which conditions get logged, which is where most manual results become incomparable.

MethodWhat it queriesWhat has to be loggedCost
Manual spreadsheetThe consumer apps, usually logged inDate, prompt, engine, pass number, named / cited, cited URL, competitors; for ChatGPT also whether it searched, which model, and whether the chat was newTime: roughly a dozen columns and a few hours a week at twenty queries
DIY on the APIsModel APIs, no consumer app contextThe same, minus the app variables, which the API does not haveAPI credits, plus the maintenance of parsers and logs
A trackerThe engines' answers without anyone's account; Claude through its APIDone by the tool; passes, cache and region are explicit5 credits × queries × engines × passes; 225 for 5 × 3 × 3x

The logged-in app, the logged-out app and the API can each return a different answer to the same prompt, because the app carries memory, location and account state. A manual sheet that mixes a colleague's logged-in ChatGPT with an incognito window is already two datasets. Querying without anyone's account keeps a colleague's history out of the data, and a competitor's name out of anyone's chat history.

Whatever tool you use, ask what it counts before you ask what it costs: whether a Claude mention and an AI Mode citation stay in separate columns, or one score is doing the separating for you.

What GA4 and server logs can and can't tell you

Google Analytics sees the clicks, not the mentions. A channel group whose session source matches this regex catches most AI referrals:

chatgpt\.com|perplexity\.ai|gemini\.google\.com|copilot\.microsoft\.com|claude\.ai

ChatGPT also appends utm_source=chatgpt.com to its links, so those are easy to isolate. Two things it will never show. A mention without a link, which on Claude is every mention, produces no session. And a click from an AI Overview or AI Mode arrives as ordinary Google organic traffic. Search Console's Generative AI performance report, open to all sites since August 2026, counts how often your pages were shown in those features, but not the clicks, the queries or what the answer said. Whether Google's AI named your brand is something only a prompt run shows.

Server logs answer a narrower question: whether the engines are reading the pages at all. OAI-SearchBot indexes for ChatGPT search and ChatGPT-User fetches a page during a live answer; Claude-User and Perplexity-User do the same for theirs, while PerplexityBot is Perplexity's indexer. A rise in the user-triggered ones on a page means it is being pulled into answers. Google-Extended is a robots.txt token, not a crawler, and has no user-agent string, so it never appears in a log and has no effect on AI Overviews either way.

What I don't have a number for

How many passes a run of five queries on four engines needs is not something the published studies answer. Żatuchin's data is three models at temperature 0.3, GPT, Gemini and Grok or Perplexity, which is not Claude through an API and not Google AI Mode; the St. Gallen runs cover four engines in four verticals, at a scale of prompts no five-line run reaches. Until someone publishes repeat data at that size, what the tool offers is what the research supports: one pass for who is in the answer today, three for whether a domain is fixed or rotating, and no claim that either is a precise rate. Where the next two years go is a guess with a basis: as ChatGPT search and Google's AI features move more answers from memory to retrieval, the cited share should rise and so should rotation between passes. That would make the stability counter more useful and the single-pass snapshot less so.

FAQ

How often should the same set be re-run?

Monthly is enough for a trend line on a stable set, and a saved project can also rerun weekly or daily and email you when share of voice moves. A rerun inside 24 hours is answered from cache. After changing a page to win a slot, wait two to four weeks before re-running the related queries: the engines need to re-fetch the page first, and an earlier pass reports the old version.

The report queries Claude through the API without web search, so every Claude answer in it is a memory answer. The consumer app can search the web and sometimes does, so a Claude mention checked by hand in the app and a Claude mention in the report are two different measurements, and only the second one is repeatable on the same terms.