Research · Dispatch #13 ·
What five AI search engines cite, and how often it changes: 22 weeks of data
24 rounds, June to November 2026: which of our glossary's terms five AI search engines cited, how often a citation came back, and what changed. With method and data.
The thirteenth GEO Glossary dispatch, and our first citation report.
Since June we have asked five AI assistants to define terms from our glossary, with one fixed question per term, and recorded whether they cite us: nine core terms in almost every round, a panel of 24 once a month, and the rest in rotation. This report covers 24 rounds, June 3 to November 2, 2026: 2,847 probes on 94 terms (2,747 checks, one per engine, term and round). The pages we measure are our own, and we wrote them. "Cited" here means cited when asked: every question ends with a request to list the web sources used, so these rates are not the rates you would see without that request.
Summary
- Five engines, five patterns. In the last eight weeks Perplexity cited our site for 33 of the 75 terms we asked about, Copilot for 23, Gemini 20, ChatGPT 14 and Claude 6, and Claude has cited none after September 21.
- Open territory gets cited. In the last eight rounds, the seven terms in our monthly panel's coined class were cited in 65% to 87% of checks on four engines; its established class was cited in 0% to 12% on all five. On the measure of our August dispatch (Dispatch #8), ChatGPT cited established terms about as often as coined ones until mid-August, and almost never since.
- A citation is a sample, not a status. When an engine had cited a term the last time it was asked, it cited it again in 68% to 75% of checks each month from June to September; when it had not, in about one in ten. October dipped to 58% (63% leaving Claude out, so Claude's drop is part of it), and November 2 was at 68%.
- The engines moved more than our pages did. From late August one engine went to zero, one narrowed and one stepped up, while we added no terms after August 26, the only two entries whose text changed were not asked about after August 31, and the edits we did make do not fit the pattern.
- Crawling is not citing. OpenAI's crawlers made about four times as many requests to our site as Perplexity's in the last eight weeks, and ChatGPT cited fewer than half as many of our terms as Perplexity did.
Since the cut (as of November 17, 2026). In the two rounds after November 2, a citation came back the next time in 16 of 32 checks (9 of 22 on November 9, then 7 of 10 on November 16), and a miss turned into a citation in 6 of 118; the June group in section 4 was cited in 9 of 30 checks, a lower share than in any earlier month; Gemini cited none of 15 terms on November 16, its fifth round without a citation; and Claude cited none of 30. Two rounds are a small sample, and every other figure in this report stays at November 2.
What we do not claim: anything about how the engines work inside; that these patterns hold for any site but ours; a recipe for getting cited.
1. Method
- The question. A fixed definitional question for each term, usually "What is X?" (one asks "Does keyword stuffing work for AI search?"), followed by "Cite the web sources you used with their URLs." The wording of a question is never edited once it is in the panel, so a change in the answers is a change in the engine or the web, not in our prompt. In June and early July seven terms were asked, in a few rounds, under an earlier wording instead; those answers count as that term's check. A set of number-seeking questions asked in two rounds for a separate test is left out of every figure here.
- Where. The normal consumer apps, not the developer APIs. ChatGPT and Perplexity are asked logged out, on the free tier; Claude, Copilot and Gemini signed in, because choosing the model needs an account, in a fresh chat. Web search is on wherever the app has a switch. All probes run from Australia.
- What counts. Our page among the sources the engine shows, inline or in its source list (a citation only in the expanded source list counts), on our own domain. A citation of our brand under a domain we do not own is not counted. A citation through another of our pages (a related entry, or our term index) counts, so "cited for a term" means the answer to that term's question cited our site, not always that term's own page.
- What we ask about. Nine core terms on every engine, in every round since July 6 and most of them before; a full panel of 24 terms once a month (June 16, July 13, August 10, September 7, October 5, November 2); the rest rotate. The panel's classes hold 7 coined terms, 6 from research papers and 11 established ones, and the nine core terms are 2, 2 and 5 (section 3 checks these classes against our territory tags).
- How we count. The unit is a term on one engine in one round (a check). "Last eight weeks" is the window from September 7 to November 2 (eight rounds). "Ever cited" is cumulative and can only rise, so it is always shown next to a window figure, never alone. The first week had three rounds (June 3 to 5) and the week of September 28 had none. Every figure uses each round's first pass: on October 26 we re-asked the misses on two engines, and this report does not count that re-check (section 8). A session that failed is re-run whole, and the re-run is the record (ChatGPT on October 12). The data stop at November 2.
- Not the Observatory's count. Our Observatory and the badges on our term pages record whether an engine has ever cited our site for that page's term, under any of our probe questions since our first probes in May, and they include citations from the October 26 re-check. They answer a different question, so their totals differ from this report's (for ChatGPT, 48 pages there and 45 terms here).
- Scripts. Every count in sections 1 to 6 is generated by committed scripts: one over the round files, one over our server logs, and one that rebuilds the history of our own edits from git. Sections 7 and 8 cite our dispatches and round notes. The methodology page and the citation probe protocol entry have the details; the Observatory shows the live series.
The probe condition stayed fixed, but the engines did not. These are the changes we could see:
| when | engine | what changed |
|---|---|---|
| throughout | Perplexity | the free tier limits searches, so panels were split across sessions; some answers came in a reduced search mode, and in some rounds from mid-September in a free preview of its advanced search, recorded per answer |
| July 20 (one round) | ChatGPT | the free tier degraded so badly that we asked most of that round signed in; it recovered and we went back to logged out from July 27 |
| from about August 31 | ChatGPT | the logged-out free tier lists about four or five sources per answer |
| August 10 | Claude | the app moved from Sonnet 4.6 to Sonnet 5 |
| October 5 (first round to see it; about September 27 on our account) | Claude | Sonnet 5.5; the app merged chat with Cowork and removed the web-search switch ("Claude searches the web when it helps") |
| October 5 | Copilot | the mode we use changed from Smart to Auto |
| about July 21 to October 26 | Gemini | the app's Flash model moved from 3.5 to 3.8, on dates we did not observe; recorded as 3.8 from October 26 |
2. Five engines, five citation patterns
Each engine cites us in its own way, and the differences are large enough to name.
| engine | full panel, Jun 16 | Jul 13 | Aug 10 | Sep 7 | Oct 5 | Nov 2 | last 8 weeks (of 75 terms) | ever (of 94) |
|---|---|---|---|---|---|---|---|---|
| Perplexity | 4 | 10 | 11 | 14 | 11 | 10 | 33 | 59 |
| Copilot | 1 | 4 | 4 | 7 | 10 | 8 | 23 | 30 |
| Gemini | 2 | 6 | 3 | 7 | 7 | 5 | 20 | 27 |
| ChatGPT | 8 | 14 | 10 | 8 | 7 | 6 | 14 | 45 |
| Claude | 4 | 6 | 11 | 4 | 0 | 0 | 6 | 20 |
Full-panel columns: terms cited out of the same 24 terms, asked in every monthly round (a monthly round also asked 6 to 20 other terms, left out of these columns). The last two columns count the distinct terms an engine cited at least once in the period.
- Perplexity is the broadest and the flattest. Since July it has cited us on 38% to 55% of its checks every month. It lists many sources on most answers, which leaves room for ours.
- Copilot rose from nothing. It cited none of our terms in the first four rounds (0 of 74 checks), then climbed month by month, from 4% of its checks in June to 24% in October; on November 2 it cited 12 of all 30 terms asked (8 of the 24 panel terms), and over the last eight weeks it cited 23 terms, second only to Perplexity. Its mode changed from Smart to Auto on October 5, but the climb had started months before. Its early climb overlaps our site's first weeks and our heaviest editing, so we do not attribute that part to the engine.
- ChatGPT narrowed. On the full panel it rose from 8 terms in June to 14 in July, then fell to 10 by August 10 and to 6 in November. From about August 31 its logged-out free tier lists only four or five sources per answer, so a page outside the top few is not shown (we see the sources it lists, not what it retrieved); its share of checks cited went from 39% in August to 24% in September. On September 7 the same question asked signed in cited our page where the free tier had not (one term, one demonstration), so the free-tier count probably understates what a signed-in user sees.
- Gemini swings. Four times it cited nothing at all across a whole round (June 3, July 20, August 3, August 24). Where a term comes from a research paper, it often cites the paper and not us. Since September it has cited between 1 and 10 of our terms per round (1 on October 26).
- Claude stopped. It cited us in every round until September 21 except a one-off status check on June 4, and on none of 112 term checks from October 5 to November 2. Dispatch #12 shows where the loss happens: our pages are still in Claude's index and open on request, but the Claude app's own search step no longer returns them for these questions.
Dispatch #4 found in June that the five engines cite largely separate sets of our pages. In the last eight weeks, of the 39 terms cited by any engine, 3 were cited by all five at some point in the window, not necessarily in the same round. A single blended "AI visibility" score would hide most of what happened.
3. What gets cited: open territory
Citations land on terms we coined and on terms backed by a research paper. Established vocabulary is rarely cited: almost never by Claude, Gemini and Copilot (0% to 5% of checks in either period), and by ChatGPT and Perplexity in 20% to 27% of checks early on, falling to 7% and 12%.
| term class | ChatGPT | Perplexity | Claude | Gemini | Copilot |
|---|---|---|---|---|---|
| coined (panel class) | 44% to 74% | 37% to 87% | 63% to 20% | 37% to 77% | 22% to 65% |
| from a research paper (panel class) | 71% to 36% | 63% to 68% | 33% to 11% | 25% to 25% | 8% to 43% |
| established (panel class) | 27% to 7% | 20% to 12% | 0% to 0% | 0% to 0% | 0% to 5% |
Share of checks in which the engine cited us, for the monthly panel's 7 coined, 6 paper and 11 established terms: the first eight panel rounds (June 3 to July 13; a one-off status check on June 4 is not a panel round), then the last eight (September 7 to November 2). Each cell rests on 24 to 58 checks; the established row has the most, because five of its terms are core terms asked every round.
The panel's classes were set when we built it. We also tag every term by the territory it sits in, and the two do not always agree: 3 of the 11 "established" terms are not established by that tag (AI search evaluation is a research concept, and two are contested), and 4 of the 7 "coined" terms are not our coinages by that tag (three are existing terms we define with our own angle, and one is a paper's concept under a shorter name). In the last eight rounds, 9 of the 14 citations of the panel's established terms were for AI search evaluation. Counted by tag instead, over every term we asked about, the result holds: in the last eight rounds the terms we coined into empty territory were cited in 18% to 86% of checks, established ones in 2% to 15%.
Dispatch #8 found in August that ChatGPT cited established terms about as often as coined ones. It did then, and it does not now. Averaging each term's own rate, as that dispatch did, gives 37% for coined and 35% for established terms over the rounds to August 17, and 49% and 3% over the last eight rounds. ChatGPT's citations of established vocabulary fell from mid-August and all but stopped from September, around the time its free tier began listing only four or five sources.
Twenty terms were never cited by any engine in any round we asked about them (2 to 7 rounds each), among them RAG, vector embeddings, knowledge graph, robots.txt, llms.txt, AI Overview and generative engine optimization itself. None of them is a term we coined. For 11 of them, the source the engines cited most often in our place was a standard, the documentation of whoever owns the term, an encyclopedia entry or the original work; for the other 9, an explainer from a large company, a search publisher or a tool vendor.
The caveat matters, and our own earlier tests narrow it. On a term we coined, ours is the originating definition, but it is rarely the only one. The coined terms that got cited in Dispatch #3 all had competing content in the same results, and in Dispatch #8 the number of competing definitions did not predict how many engines cited a coined term. Dispatch #3 also found that Gemini's high share of coined-term citations followed our Google rank, not our being the originator. So the table shows that open territory, where no standard or vendor owns the definition, gets cited; it does not show that coining a word earns citations by itself. The one established term we ran as a matched control against a coined one (E-E-A-T, paired with cite-ability) was cited once in 80 checks. We retired that arm from October 19, with two other parts of the panel, under a standing rule: at our five-month review, any part that, in every round to October 12, had detected nothing the core panel did not was dropped.
4. Citations churn: one probe is not a status
A citation varies in four separate ways, and a single "cited / not cited" reading captures only the first.
- Inclusion: cited one week, not the next, for the same question.
- Prominence: cited both times, but first one week and fifth the next, or through our entry one week and through a related page the next.
- The rendered address: cited both times, but under the wrong domain once (section 7).
- The probe condition: a changed tier or search mode on the engine's side (section 1).
The inclusion axis is the one we can count. Each check below is compared with the previous time the same engine was asked about the same term: if it cited the term then, how often did it cite it again, and if it did not, how often did it cite it now?
| month | checks after a citation | cited again | share | leaving out Claude | checks after a miss | cited | share |
|---|---|---|---|---|---|---|---|
| June | 62 | 45 | 73% | 71% | 368 | 36 | 10% |
| July | 94 | 64 | 68% | 66% | 379 | 54 | 14% |
| August | 134 | 100 | 75% | 74% | 361 | 47 | 13% |
| September | 96 | 67 | 70% | 70% | 239 | 21 | 9% |
| October | 98 | 57 | 58% | 63% | 296 | 28 | 9% |
| November 2 | 41 | 28 | 68% | 72% | 109 | 13 | 12% |
A check is one engine asked about one term in one round. June includes re-checks one or two days apart in our first week; without them, June is still 73%.
From June to September about seven citations in ten came back the next time, and in every month about one miss in ten (9% to 14%) turned into a citation. October dipped to 58%, and to 63% leaving Claude out, so Claude's drop is part of it but not all; November 2 was back at 68%. A fixed group shows no trend either: the 15 pairs of an engine and a core term that were cited in June (leaving out Claude) were cited in 53% to 67% of their checks in each later month to November 2.
A count that looks further back, at whether the engine had cited the term in any earlier round, falls from 66% in June to 40% in October. Through September that fall comes from what is being counted, not from the engines: the set of terms ever cited only grows, so terms cited once in July are still in it later, and it fell from 66% to 55% while the count above held at 68% to 75%. October's further drop shows in both. Counted across engines in the same way (a term cited again by any engine, after any engine had cited it), it falls from 90% to 49%. The data file has all of these series.
For a publisher, the practical reading is that one check is a sample. A citation had about a seven-in-ten chance of showing up again the next time, and a miss about one in ten of turning around. A term cited once is not "cited", and a term missed once is not "lost": ask again in later rounds, count the misses as well as the hits, and compare each check with the one before it, not with every citation since the start.
5. The engines change more than the pages do
The clearest movements in the period came from the engines, not from our pages. Claude's drop and Copilot's step-up each line up in time with a change to that engine's product that we could see (section 1), and ChatGPT's narrowing is itself one. Lining up in time is not proof that the product change caused the movement.
| event | from | what moved |
|---|---|---|
| Claude stops citing us | about September 27 | 0 of 112 checks from October 5 to November 2, while the other four engines kept citing 5 of the 6 terms Claude had cited since September 7 (the sixth was not asked again) |
| ChatGPT's free tier narrows its source list | about August 31 | about four or five sources per answer; share of checks cited 39% in August, 24% in September; full-panel terms 10 on August 10, 8 on September 7, 6 on November 2; established terms all but stop from September (section 3) |
| Copilot steps up on the monthly full rounds | October 5 | 12 of all 32 terms asked that round (8 of 32 on September 7), and 12 of 30 on November 2; of the 24 panel terms, 7 on September 7, 10 on October 5 and 8 on November 2; the weekly rounds between October 5 and November 2 stayed at 1 to 4; its mode changed from Smart to Auto that day |
That our pages changed less needs stating with care, because they were not frozen:
- We added no new term after August 26.
- After August 26 the text of only two entries changed: our cited-version lag entry (August 28, September 10 and 21, October 22) and our Claude citation entry (October 29). Neither was among the terms we asked about from August 31 to November 2.
- Three dispatches linked to terms we ask about: Dispatch #9 on August 29, #11 on September 23 and #12 on October 28.
- Each round sets the checked date on every page it asks about. Where an engine cites a page for the first time, the round also upgrades that page's badge and adds a line to its change log.
- On October 22 the modified dates of 21 entries were moved forward to each entry's last real edit (May to June 21), with no change to their text. That changed the dates search engines see and can prompt a recrawl.
Several of these come before an event: the edits to the cited-version lag entry and the links from Dispatches #9 and #11 fall before ChatGPT's narrowing or Claude's drop, and the checked dates change every round. What they do not fit is the pattern. Each touched pages that all five engines were asked about, yet at Claude's drop one engine stopped and four kept citing the same terms. The October 22 date change came after Claude's drop, after ChatGPT's narrowing and after Copilot's 12 on October 5. A publisher who reads only their own analytics sees a drop and cannot tell an engine's change from their own content decaying. A fixed panel asked every week can: when one engine moves and four do not, on the same page, the cause is specific to that engine (its product, its search step, or what its search returns), not a change on our site as a whole. Which of those it was, the panel cannot tell.
6. Crawling is not citing, and referrals are a trickle
How much a company's bots fetch from a site says little about how often its engine cites it.
| company | crawler requests, last 8 weeks | answer-time fetches, last 8 weeks | its engine's terms cited, last 8 weeks (of 75) |
|---|---|---|---|
| OpenAI | 4,369 | 625 | ChatGPT: 14 |
| Perplexity | 1,035 | 0 | Perplexity: 33 |
| Anthropic | 709 | 91 | Claude: 6 |
| Microsoft | 1,353 | none logged | Copilot: 23 |
| 1,207 | none logged | Gemini: 20 |
Rows in our AI-crawler log. Answer-time fetches are the assistants' own fetchers (ChatGPT-User, Claude-User, Perplexity-User) collecting a page while writing an answer, some of them while answering our own checks; every other listed bot counts as a crawler. Google's row counts the Googlebot and GoogleOther user agents and Microsoft's counts bingbot; both serve far more than Gemini and Copilot, so only the first three rows compare.
OpenAI's crawlers made about four times as many requests as Perplexity's, and ChatGPT cited fewer than half as many of our terms. Anthropic's bots made 800 requests in the same eight weeks, 325 of them in the four weeks after Claude's drop (September 28 to October 25) and 10 in the last week, while Claude cited 6 terms, none after September 21. Crawl volume is not a proxy for citations; with three comparable companies and one site, we claim no more than that.
Visits that arrived from an AI assistant (its address as the referrer) are few: 51 in five months (16 in June, 14 in July, 9 in August, 9 in September, 3 in October), of which 7 came from Australia, where our own checks run. This is the publisher's measurement problem in AI search, not our success measure: the answer is read in the assistant, and a visit is the only part of it a site's own analytics can see. Counting citations directly is the only way we have found to see what the engines do with a site.
7. When an engine cites you wrong
"Cited" is not the same as cited correctly. We saw three kinds of failure in our own data, each invisible to a counter that records only cited or not.
| failure | what we saw |
|---|---|
| a stale copy | Claude cited our page's live address while giving figures we had corrected on June 21, in every weekly round from June 29 to September 14 apart from one run. The figures alone could not show a stored copy; the wording of at least three answers matched a version of the page from late May, which points to one (Dispatch #11, corrected October 22) |
| the wrong address | Copilot rendered our brand under two domains we do not own, one that did not resolve and one that was parked when we checked: five citations on July 13, with our exact page path, and one on August 24 |
| a source readers cannot open | five programmatic pages from one GEO vendor, cited by Claude eight times on our terms from October 5 to November 2 (seven in Dispatch #12, one more on November 2); on October 20 and 27 they returned "Page not found" to requests sent under a browser's user-agent name |
Dispatch #9 sets out the three checks a citation has to pass: the source supports the claim, the citation points to the right source at an address that resolves, and the version cited is current. The three failures above fail the second and third checks.
8. Our own corrections in this period
We make mistakes in this work too, and we correct them in public. All corrections to our entries are listed on the corrections and fact-checks page and its archive.
- A finding that did not survive its own test (June 17). Dispatch #3 reported on June 13 that Gemini credited the originator of a term, and promised to test that against Google rank. The June 16 round ran the test: Gemini's citations followed our rank, not our authorship. We rewrote the dispatch on June 17 and 18, but the page did not say so in a dated note until November 6.
- The wrong column of a table (June 21). Fifteen entries that quote the GEO paper's Table 1 had cited the plain word-count sub-column as the position-adjusted metric. We corrected them to the paper's position-adjusted "Overall" column, for example Quotation Addition at 27.2 against a baseline of 19.3.
- A stale-copy finding with the wrong date and a failed test (October 22). Dispatch #11 put the correction behind a stale-copy episode on August 4 instead of June 21, and its test for a stale copy failed because our own change log and a footnote still carried the old figures. We corrected the dispatch, withdrew its 48-day duration and its claim that the lag ended on its own, corrected Dispatch #9 and the cited-version lag entry the same day, and now date a stale answer by wording that changed, not by figures our correction record repeats.
- A headline that could only rise (October 14). Our monthly review headlined the share of terms ever cited. That number cannot fall, and on the monthly full panel it hid a decline from 19 of 24 terms cited in July to 13 in October (as the panel is counted now). Every cumulative figure in this report now sits next to a window figure.
- A re-check that leaned one way (October 26). In that round we re-asked only the checks that had not cited us, on two engines, which can only add citations. This report counts the round's first pass, and the data file carries the record next to it (first pass against the record: ChatGPT 1 term against 3, Gemini 1 against 5, one of the five a partial citation). From November 2, a check is re-run only when a session fails, and then the whole session is re-run.
One disclosure about our site: the citation badges on our term pages only ever upgrade, because a single "not cited" reading is too noisy to remove a recorded citation. A badge records that an engine has cited us for that page's term at least once, not that it still does. This report is where the current picture lives. Since November 18, 2026, a cited badge is shown hollow, as "previously cited", once that engine has not cited us in its last four checks of the term (a partial result counts as a miss); the record behind it still only moves up.
9. Limitations
- One site, and we are both the optimizer and the measurer. These are our pages' citations, not a general rate for the engines.
- The panel leans toward our own terms. Coined and paper terms are 13 of the panel's 24 terms and 4 of the 9 core terms. By our territory tags the lean is smaller but clear: 10 of the 24 panel terms are coined into empty territory or come from papers, against about a quarter of the glossary. They are cited far more often than established ones, so totals such as "33 of 75" depend on that mix; the per-class rates in section 3 do not.
- Our own links move what we measure. Our dispatches and internal links can change how the engines find our pages; we recover every edit and date change from git history, and section 5 lists those after August 26.
- Definitional questions, with one exception. "What is X?" is one kind of question. How-to and comparison questions may be cited differently.
- Engine versions are mostly hidden. On consumer apps we see the model name only where the app shows it, and the free tiers do not.
- One region, and one condition per engine at a time. All probes run from Australia, and no probe varied the region; Dispatch #12 ruled region out only as the cause of Claude's drop.
- Small numbers. A single engine on a single round covers 13 to 44 terms. We report counts per engine and do not fit a model pooled across engines.
10. Data, method and contact
- Data: two CSV files, counts only, released under CC BY 4.0. Per-term detail and the engines' answer text are not published.
- By round, engine and term class: the terms asked about, the terms cited (first pass), and the terms as recorded after the October 26 re-check. It gives the per-round, per-month and per-class figures in sections 2, 3 and 8.
- Counts across rounds: the full panel per engine; for section 4, the checks cited again after a citation and after a miss, the June cohort, and the series that look back at any earlier citation, by the same engine and by any engine; and the distinct terms cited in the last eight weeks and ever, per engine.
- Method: the methodology page, the citation probe protocol and the Observatory.
- Contact: if you want the data in more detail, the method, or a run on your own terms, write to hello@aisearchglossary.com.
Data: 24 probe rounds, June 3 to November 2, 2026 (rounds after November 2 are not included), and our server logs from June 8 to November 2. The counts in sections 1 to 6 are generated by committed scripts from the round files, the logs and our git history.
More dispatches
- Dispatch #12Claude stopped citing us in late September. Four other engines kept citing the same pages.
- Dispatch #11We corrected a number; an AI engine kept quoting the old one, and our first test of why was wrong
- Dispatch #9A citation is not proof: three ways AI search cites you and still gets it wrong
- Dispatch #8We built this glossary on coining our own terms. Does AI search actually reward it?
- Dispatch #7We refuse to invent benchmarks. Does AI search punish us for it?
- Dispatch #5GEO's most-cited numbers, checked against the papers they come from
- Dispatch #4One panel, five engines, mostly separate citation sets: 'cited by AI' is not one thing
- Dispatch #3Cited more on Gemini, less on ChatGPT: a gradient, and what it is not
- Dispatch #2Google caught up: the AI-citation gap looks like a reporting lag
- Dispatch #1AI engines cited this page before Google indexed it