How stable are AI citations? Not very, and the instability is built in. Ask an AI engine the same question twice, minutes apart, and it will often cite different pages, though nothing on the web has changed in between.
In independent studies published in 2026, two runs of one question on the same day had only 15% to 66% of their cited sources in common, depending on the engine, and the overlap kept shrinking for weeks afterwards.
So a single check proves little. What holds still is the rate: how often a company is cited across many runs, days and wordings, which is the number AI search visibility work should be judged on.
What “stable” means for an AI citation
A citation can move in six ways, and most arguments about whether AI citations are stable mix them up. Each one needs handling on its own when you measure.
- Between runs. The same question, asked twice within minutes, cites different pages.
- Over time. A page cited all month drops out, or a new one arrives, as days turn into weeks.
- Between wordings. “What is” and “what’s” in the same question bring back different sources.
- Between devices and places. A phone and a laptop, or two cities, see different sources for one question.
- Between engines. ChatGPT, Gemini and Google’s AI Overviews cite different pages for the same question, because each engine runs its own searches and picks from what they return.
- No citation at all. The answer cites nobody, because the engine answered from what it already knew. That is not a lost citation, and counting it as one distorts every figure.
How much AI citations change, engine by engine
Google’s AI Overviews
The strongest evidence on Google comes from a study led by the New Jersey Institute of Technology, peer reviewed and presented at the ACM’s SIGIR conference in July 2026. It compared Google’s results, AI Overviews and Gemini on 11,500 queries collected in December 2025.
To test consistency, the researchers ran 100 of the queries twice at the same moment. On average, 66% of the sources the two AI Overviews cited between them appeared in both, against 78% for Google’s ordinary results and 46% for Gemini.
Switching between a phone and a desktop cut the AI Overview figure to 55%. Google’s ordinary results moved too, from 78% to 69%, but less.
Over months the gap widens. Researchers at Ruhr University Bochum and the Max Planck Institute for Software Systems repeated 4,706 queries about two months apart in 2025, in a paper revised in May 2026.
Only 18% of the pages AI Overviews cited were common to both runs, against 45% for Google’s organic results.
ChatGPT, Gemini and Perplexity
A University of Toronto study published in October 2026 asked 651 questions about a US internet provider every day for 90 days, through each engine’s public API. Between two runs on the same day, ChatGPT’s cited pages overlapped by just 15%.
Over weeks, the researchers measured how much of the pattern of which pages got cited, comparing pages on the same site, carried over from one fortnight to the next. Even within a single fortnight, only about two thirds of it repeated.
Two weeks on, 40% of ChatGPT’s pattern still held, 47% of Gemini’s and 22% of Perplexity’s. Ten weeks on, it was 14%, 29% and none at all.
The studies measure in different ways, so read the figures as a direction, not a league table. All of them point the same way: AI answers change their sources more than Google’s ordinary results where both were measured, and less of today’s list survives each week.
Why AI answers cite different sources each time
The question becomes new searches every time
An AI engine does not look up your question as typed. OpenAI’s help page on ChatGPT search, updated in October 2026, says it “typically rewrites your query into one or more targeted queries” for its search providers.
The Toronto researchers watched it happen. ChatGPT split each question into about four keyword searches, and added the current year to 23% of them and a brand the user never named to 28%. Between two runs on the same day, its searches overlapped by 2.7%.
Google works the same way. Google’s guide to its AI features says AI Overviews and AI Mode can run many related searches before answering, and that the two “may use different models and techniques, so the set of responses and links they show will vary”.
The wording changes the searches
Because the searches are built from the words asked, small edits matter. In the New Jersey study, writing “what’s” for “what is”, “U.S.” for “United States” or adding a question mark cut the similarity of AI Overview sources by 29%. Google’s ordinary results lost 14%.
Every answer is written fresh
The model then writes a new answer from what it found, and that step has its own randomness. In the Bochum and Max Planck study, even with the models’ randomness setting at zero, 9% to 27% of yes or no questions flipped their answer within five minutes.
The web and the models keep moving
Over weeks, pages change, indexes refresh and models are replaced. Bing, launching new AI citation reports in June 2026, wrote that citation patterns “can shift due to changes in user behavior, evolving models, freshness signals, partner refresh cycles, and broader changes across the web itself”.
Sometimes there is no search at all
Each engine decides for itself whether to search. OpenAI says ChatGPT may search automatically when a question “would benefit from current information”, and Google says AI Overviews often do not appear. Through its API, Perplexity cited no source at all in 96% of the Toronto study’s answers.
Which AI citations tend to stay
The research on what keeps a citation is thinner than the research on how much citations move. A 2026 Sciences Po review of 45 studies found no technique with a stable, lasting effect across engines.
Ranking well in Google is no guarantee either. In the New Jersey study, AI Overviews and Google’s own ordinary results had fewer than a fifth of their sources in common, a gap that shows why ranking and being cited are now separate jobs.
What the evidence does support is narrower. In the Toronto data, every page ChatGPT cited was one its own searches had found, so being the clearest answer to the searches an engine runs is the closest thing to a lever.
Pages whose titles and paragraphs matched the questions people ask were cited more than other pages on the same site.
Brands behave differently from links. The same study found that ChatGPT writes brands it already knows into its searches and names brands with no page behind them, so a company can stay named in an answer after the link to its site has gone. Track the two separately.
For a small B2B vendor, the way into answers runs mostly through other people’s pages. In the study’s second market, small B2B software vendors, a vendor was named mainly when a cited third-party page named it, which is where coverage and links from trade sites earn their keep.
How to measure AI citations when they keep moving
Treat a citation as something you sample, not something you look up. The Sciences Po review, published in July 2026, concludes that a sound measurement has to vary along at least four lines: the run, the wording, the date and the engine.
Choose the questions
Take 20 to 30 questions buyers ask before they shortlist anyone, from sales calls, enquiry forms and the queries in Search Console. Split them into brand questions, which name you, and category questions, which do not. A strong result on one says nothing about the other.
Write each question three ways, in its usual wording and two rewordings, since the wording alone changes the sources.
Fix the conditions
Use a logged-out browser or a clean profile, the same type of device and the same location every time, and note all three. Each of them can change the sources, so changing one between checks makes the checks impossible to compare.
Record every run
Run each question several times per check, on each engine you care about. For every run, record whether you were cited and which page, whether you were named without a link, and whether the answer cited anyone at all.
Keep the answers that cite nobody in the count and report them on their own line. The Sciences Po review warns that dropping them biases the result, and on some engines they are most of the answers.
One row per question, wording and engine is enough. The rows below are made-up examples: a week of runs on ChatGPT, 14 per question. “Named only” counts runs that named you without citing your site.
| Question | Type | Cited | Named only | No citations | Rate |
|---|---|---|---|---|---|
| What does an outsourced SDR cost? | Category | 3 | 2 | 1 | 21% |
| How much is an outsourced SDR? | Category, reworded | 1 | 2 | 0 | 7% |
| Is [your company] a fit for mid-market software firms? | Brand | 9 | 3 | 2 | 64% |
Read a rate, with its margin
Your citation rate is the share of runs that cite you. Like any sample, it carries a margin of error, roughly 1.96 × √(p × (1 − p) ÷ n) either side, where p is the rate and n the number of runs.
That is plain arithmetic, not a finding, and it assumes each run is independent, which repeated questions only roughly are. At a 50% rate it works out as follows.
| Runs | Margin either side |
|---|---|
| 5 runs | 44 points |
| 10 runs | 31 points |
| 20 runs | 22 points |
| 30 runs | 18 points |
| 50 runs | 14 points |
| 280 runs | 6 points |
So one question needs dozens of runs before a change means anything. Pool questions instead: 20 questions run twice a day for a week give 280 runs, enough to trust a move of ten points or more from one week to the next.
The Sciences Po review gives the same advice in its own terms: keep repeating the measurement until the margin is narrow enough for the decision you need to make.
Let the free reports run alongside
Google and Bing now count AI citations on their own engines. Search Console’s generative AI reports show your impressions in AI Overviews and AI Mode, and have covered every site since 31 August 2026.
Bing Webmaster Tools counts your citations in Copilot and Bing’s AI answers, page by page. Since June 2026 it also shows Citation Share, your percentage of all citations for a query, and Compare, which can lay the last 30 days over the 30 before.
A citation rate shows presence, not pipeline. Read it beside the signals that show up when buyers read an answer and never click, such as branded searches and a “how did you hear about us?” question on the enquiry form.
When a lost citation is worth acting on
A citation missing from one run is noise, and so is one that comes back. Act when the rate for a group of questions falls across a full fortnight, on more than one question or engine. Then work through the causes in this order.
- Did the answer cite anyone? If the engine stopped searching for that question, nobody lost a place.
- Can the engine still use the page? Google says a page must be indexed and eligible to show with a snippet to appear in AI Overviews or AI Mode. Check the page in Search Console, and the settings that keep a site out of AI answers.
- What replaced you? Open the pages now cited. If they answer the question more directly or more recently, that is the fix.
- Did the searches change? Bing’s AI Performance report lists the phrases the AI searched with when it cited you. If they shifted, the page may now match the wrong searches.
Mistakes when reading AI citation data
- Reporting a screenshot. One run, pasted into a board deck, shows what happened once. The next run may say the opposite.
- One number for every engine. The engines search differently and cite different numbers of pages, so a blended rate hides where you are strong and where you are missing.
- Dropping empty answers. Leave out the answers that cite nobody and your rate looks better than a buyer’s experience of it.
- Crediting an edit for a comeback. Citations drop and return on their own. Judge a page change on the rate over the next fortnight, not on the next run.
- Tracking only brand questions. Being cited when a buyer names you is the easy case. Category questions, asked before anyone has a shortlist, are where new pipeline starts.
● About the author
Hamza Hashim is the marketing strategist at Katama. He owns the channel mix and the order it gets built in: what ships first, what waits for evidence, and what gets stopped when the evidence says stop. Meet the team.