How Stable Are AI Citations? 10 Things Published Research Tells Us

Short answer: not very. Across every study we found, the sources an AI search tool cites for a given question change a lot from one answer to the next, from one day to the next, from one engine to the next and even between settings inside the same product. A small core of sources and brands tends to persist, while a large tail rotates. That has real consequences for anyone trying to track, improve or sell “AI visibility”.

A note on what this article is and is not. The title you may have seen elsewhere promises findings from tracking a large set of citations ourselves. We did not run that study, so this guide does not claim one. Instead, it collects what published studies and a recent academic survey report about citation stability, says who ran each one and how large it was, and flags the limits. Most of the studies come from companies that sell AI-visibility tracking or services, and several are small. Treat them as directional evidence, not settled fact.

This guide is written for adult business owners, directory operators and platform marketers who are being offered “AI SEO” and need to know what is measurable. It is non-graphic and general.

How we read the research

We read nine sources in full or in part: seven company or blog studies with named methods, one academic paper (an arXiv preprint) and one critical survey of 45 studies. We note each study’s sample, dates and conflicts of interest. We did not re-run any of them, and we found no study focused on adult-industry queries. Numbers below are as the authors report them. Where a figure comes from a secondary summary, we say so.

1. Exact repeats are rare

If you ask the same question twice, you almost never get the same set of cited sources.

• EWR Digital’s study, published 29 August 2026 and covering 12,778 answers and 118,050 citations across five engines from 13 January to 25 June 2026, found that ChatGPT returned an identical citation set on repeat only 0.06% of the time, and no engine exceeded 2.5%.
• A 28-day study by Glen Allsopp of Detailed, published 23 September 2026, checked more than 1,300 prompts a day on ChatGPT and Google AI Overviews. For two randomly chosen days, an identical brand list appeared 0.3% of the time on ChatGPT and 1.1% on AI Overviews.

What it means: do not treat one answer as “the” answer. A single check tells you very little.

2. Typical overlap between answers is low

Researchers often score overlap with the Jaccard index, where 1.0 means identical sets and 0.0 means nothing in common.

• EWR reported median overlap for repeated asks of 0.133 on ChatGPT, 0.231 on Gemini, 0.333 on Perplexity and DeepSeek and 0.417 on Claude. Every engine was below 0.45.
• OrganiKPI’s August 2026 test ran 28 questions four times on one day through a scraper on ChatGPT. Two runs of the same question overlapped by 0.39 on average. In that test, 55.9% of question-source pairs appeared in only one run and 11.8% appeared in all four.
• Parse’s study of 393,201 citation links found that consecutive answers shared 33.3% of hostnames on ChatGPT Search and 24.7% on Google AI Mode, with a gap of about two days.
• The critical survey of 45 studies cites reports of daily source-level Jaccard scores of about 0.34 to 0.42.

The studies use different methods, so the numbers do not line up neatly. Some compare repeats on the same day, others compare days or months, some count domains and others count full URLs or brand names, and the prompt sets differ. The direction is consistent, though: roughly a quarter to a half of cited sources carry over, and exact repeats are the exception.

3. Stability differs a lot by engine

GetMentions AI published a study on 9 July 2026 covering seven consecutive days in June 2026: 2,398 queries from 56 brand accounts, 67,144 answers and 530,875 citations across ChatGPT, Google AI Mode, Perplexity and Gemini. It reported that about 69% of a typical answer’s sources change from one day to the next, with daily churn of 88.3% on Gemini, 79.2% on ChatGPT, 75.9% on Google AI Mode and 44.4% on Perplexity. Only 0.4% of Gemini sources and 1.1% of ChatGPT sources were cited on all seven days, against 11.1% on Perplexity.

The authors call their results directional, because they come from one tracking platform and one seven-day window.

What it means: a result that holds for one engine says little about another. Track engines separately.

4. There is a stable core and a volatile tail

Instability is not uniform. The Detailed study found that the brand named on the most days appeared on 82% of days for ChatGPT and 89% for AI Overviews. Brands named on at least 80% of days (the “core”) changed by 13% day to day on ChatGPT and 11% on AI Overviews, while the tail changed by 78% and 65%. Core cited domains changed by 14% and 12%, against 74% and 61% for the tail.

Parse’s historical ChatGPT panel, with about 22 runs per question, found about 81.3 distinct domains per question, but only 1.4 domains appeared in at least 80% of answers, and those accounted for 22.1% of citations.

What it means: some sources do hold a position, but they are few. For most brands the realistic goal is a higher probability of appearing, not a guaranteed slot.

5. Different engines cite different sources

• GetMentions found that 84% of domains cited for a query came from a single engine, and that any two engines shared about 4.3% of sources on a single day and 8.3% over the week.
• Detailed found that ChatGPT and AI Overviews agreed on the most-present brand only 27% of the time.
• The survey reports URL-level Jaccard similarities of 0.11 to 0.18 among organic Google, AI Overviews and Gemini, and concludes that visibility is indexed by engine and surface.

What it means: “AI visibility” is not one number. Reporting a single score across engines hides large differences.

6. Settings inside one product change the sources

Semrush, working with Kevin Indig and published 30 June 2026, ran 100 prompts twice on GPT-5.2, once at minimal reasoning and once at high reasoning. Only 25.6% of cited domains were shared between the modes. The citation rate rose from 50% to 68%, citations per response from 2.6 to 4.5, and the high-reasoning mode ran 4.6 times more sub-queries.

This is a small study on one model, run to promote Semrush’s own toolkit, and the page does not explain how prompts were chosen. Still, it points at a simple fact: the same prompt can behave differently depending on the mode.

What it means: when you track, record which mode and interface you used. Comparing runs from different settings compares different things.

7. AI sources only partly overlap with Google’s organic results

A paper by Kirsten and colleagues at Ruhr University Bochum and the Max Planck Institute for Software Systems compared Google’s organic results with five generative systems across 4,706 queries. It reported that fewer than 50% of Google AI Overviews links match the top-10 organic results, and that on average 53% of the domains AI Overviews cites are not in the organic top 10 (27% are not in the top 100). Overlap was lower for some other systems, and the paper’s text is inconsistent about Gemini, so we do not repeat a figure for it. We read the arXiv version, whose venue we could not confirm.

The critical survey also cites a report that page overlap across two months was 18% for AI Overviews, against 45% for organic Google. We rely on the survey for that number and did not check the original.

What it means: ranking well in organic search does not guarantee a citation, and a citation does not need a top-10 ranking. Both can move independently.

8. Month-to-month change is hard to tell from noise

This is the most useful finding for anyone reporting results.

• EWR found that overlap between consecutive months ranged from 0.318 to 0.458, close to the 0.309 seen when the same prompt was simply asked again. The author concludes that month-over-month change is hard to distinguish from noise. EWR also reports that a single run has a 95% confidence interval of about plus or minus 33 percentage points, narrowing to about plus or minus 10 points after seven runs. One correction was added after publication, and the author sells the service the study supports.
• OrganiKPI treated swings smaller than about 38% of a question’s mean credited-source count as noise.
• The critical survey suggests seven to eight repetitions per prompt as a starting point, and says that is not a standard.

What it means: a report that says “visibility up 12% this month” from one run per prompt is probably reporting noise. Ask how many repetitions were used.

9. What tends to get cited

Several studies note patterns, but they are associations, not causes, and the samples are small or tied to one platform.

• OrganiKPI found that only 1 of 238 credited URLs (0.4%) was a homepage, and that credited sources were a median of one folder deeper than dropped ones. Among sources with a visible date, 69.6% of credited ones were from 2026 or later, against 62.0% of dropped ones.
• GetMentions found that editorial and business pages made up 64.9% to 81.3% of citations, depending on engine, and that Reddit was the most-cited domain on ChatGPT and Gemini while YouTube led on Google AI Mode and Perplexity.
• The Bochum paper found that GPT-based systems lean on corporate entities and encyclopedias and rarely cite social media or forums, while organic search draws up to 35% of results from social media and forums.

The survey’s conclusion is blunt: no reviewed technique shows a stable, longitudinal, cross-platform causal effect on organic discoverability, and generic tricks such as keyword stuffing transfer poorly across engines.

What it means: clear, specific, up-to-date pages on your own site are a sound bet. There is no proven formula.

10. The measuring itself is shaky

Even when sources seem to drop out, the cause may be measurement.

• The Detailed study found that for 55 brands, 36% of apparent drop-offs on ChatGPT and 30% on AI Overviews were the same company under a different name.
• The survey cites a study in which 57.8% of ChatGPT repetitions did not activate web search, and warns that excluding answers without citations introduces selection bias.
• AI Overviews do not appear for every query. The survey cites one report that overall activation was 13.7%, rising to 64.7% for question-phrased queries.
• A secondary summary of an October paper says citations visible in ChatGPT’s interface only partly match the sources returned through OpenAI’s API. We could not read the original paper, so treat this as a caution, not a measured figure.
• The survey also notes that citations are not proof that a source supports a claim, and that citation metrics do not predict clicks, conversions or revenue.

What it means: a tracking number is only as good as its method. Prefer tools and reports that disclose prompts, interface, dates, repetitions and how names are matched.

What this means for adult businesses

We found no study focused on adult-industry queries, so we cannot say how stable citations are for them. Anything specific would be a guess.

Here is what the general evidence suggests, labelled as our own reading.

• Do not buy “guaranteed AI citations”. The evidence shows that citations vary run to run, and Google says nobody can guarantee a number one ranking.
• Google says AI features need no special optimisation. Its AI features guidance says AI Overviews and AI Mode need no extra work, that standard SEO practice applies, and that a page needs to be indexed and eligible to appear with a snippet. It recommends allowing crawling, using internal links, keeping important content in text, matching structured data to visible text and keeping Merchant Center and Business Profile information current. It says there is no need for new machine-readable files, AI-specific text files or special markup, and that AI feature traffic appears in the Search Console Performance report under Web search type.
• Crawler access is the first thing to fix. If you use an age gate, Google says to allow verified Googlebot through, using a reverse and forward DNS check or its published IP ranges. Pages that Google cannot reach cannot be cited in Google’s features.
• Classification still matters. Google’s explicit-content guidance describes grouping explicit content on a separate domain or subdomain. How AI features treat explicit queries is not covered in the pages we read, so we make no claim.
• Consent and trust records matter more. They protect you with search, payments and platforms, whichever way citations move.
• Keep content specific, current and non-explicit where the goal is a commercial business site.

How to track AI citations responsibly

If you want to monitor AI citations, the research points to a short list of good practices. These are our summary of what the studies imply, not a standard.

1. Pick a defined prompt set that reflects real customer questions, and keep it fixed.
2. Run each prompt several times. The survey suggests seven to eight as a starting point.
3. Record the engine, mode, location, logged-in state and date for each run.
4. Track engines separately.
5. Normalise brand and domain names, so renamed or variant forms do not look like drop-offs.
6. Report ranges, not single numbers, and compare periods using the same method.
7. Separate “mentioned” from “cited” from “clicked”. A mention is not a visit.
8. Check against Search Console. Google says AI feature traffic appears in the Performance report under Web search type.
9. Keep a change log of your own site changes, so you can look for associations.

Hiring an AI SEO specialist

Some businesses will want help. If you decide to hire an AI SEO specialist, apply the same skepticism you would to any SEO partner, plus a few specifics. Google’s guidance on hiring an SEO says no one can guarantee a number one ranking, lists unsolicited emails and claims of a special relationship with Google as red flags and suggests read-only Search Console access to start.

What a good AI SEO specialist should do

• Fix the basics first: crawl access (including behind an age gate), indexing, clear page structure and honest structured data.
• Improve content so that it answers real questions specifically and accurately, with named authorship and sources where appropriate.
• Set up measurement with disclosed methods and ranges.
• Explain what is within your control and what is not.

Questions to ask

1. How many times do you run each prompt, and on which engines and modes?
2. Can you show confidence ranges, not single scores?
3. How do you handle name variants and renamed brands?
4. How do you separate noise from real change?
5. What do you change on our site, and how will we see the effect in Search Console?
6. Will you guarantee citations or rankings? (A good answer is no.)

Red flags

• Promises of guaranteed citations in a named AI product.
• One screenshot offered as proof.
• “Special files” or markup sold as the secret, which Google’s AI features guidance says are not needed.
• Reports that cannot be reproduced.

Frequently asked questions

How stable are AI citations?

Not very. Published studies report low overlap between repeated answers, high daily churn on several engines and a small stable core, with a larger rotating tail.

Which AI engine is most stable?

In GetMentions’ seven-day study, Perplexity had the lowest daily churn (44.4%) and Gemini the highest (88.3%). That is one vendor’s study in one week, so check recent data before relying on it.

Is one check enough to know if I am cited?

No. The research suggests many repetitions per prompt to separate signal from noise.

Do I need special files or markup for AI search?

Google says no: it says AI Overviews and AI Mode need no special optimisation or AI-specific files.

Does ranking in Google guarantee AI citations?

No. Studies report that a large share of AI-cited domains are outside the organic top 10.

Are these studies reliable?

Treat them as directional. Many are run by companies that sell tracking or services, several are small and most are not peer reviewed. The academic survey we cite is a single-author preprint.

Is there data for adult-industry queries?

We found none.

Next steps

If you track AI citations, track them like a scientist: fixed prompts, many runs, ranges and honest caveats. If you optimise for them, start with the basics Google describes, which are the same as for search generally.

Escort Marketing Agency (EMA) is a B2B digital marketing agency that works with adult businesses on technical SEO, website development and online visibility. If you want to hire an AI SEO specialist who will measure honestly and promise nothing they cannot deliver, Connect with us today.

This article is general information and not legal advice. Check the rules for each market you serve.

References

• Detailed (Glen Allsopp), “How Volatile Are AI Responses? (70K Answer, 28 Day Study)” (23 September 2026): https://detailed.com/ai-volatility/
• GetMentions AI (Anirudh Agarwal), “AI Citation Volatility: A 530,875-Citation Study” (9 July 2026): https://www.getmentions.ai/blog/ai-citation-volatility-study
• Semrush, “Only 25% of cited sources overlap between ChatGPT’s different reasoning modes” (30 June 2026): https://www.semrush.com/blog/chatgpt-reasoning-ai-visibility/
• EWR Digital (Matthew Bertram), “We Ran 12,778 AI Answers. ChatGPT Repeated Itself 6 Times in 10,000.” (29 August 2026; corrected 5 September 2026): https://www.ewrdigital.com/ai-answer-variance
• OrganiKPI (Daniel Shashko), “ChatGPT Citation Variance: 4 Runs, Same 28 Questions” (16 August 2026; updated 2 October 2026): https://organikpi.com/blog/geo-ai-search/chatgpt-citation-variance-study/
• Parse (Dimitry Apollonsky), “AI citation volatility by industry” (8 July 2026; updated 12 October 2026): https://parse.gl/research/ai-citation-volatility-by-industry
• Kirsten et al., “Characterizing Web Search in the Age of Generative AI” (arXiv preprint): https://arxiv.org/html/2510.11560v2
• Olivier Martinez, “Optimizing Visibility in Generative Engines: A Critical Survey of Generative Engine Optimization (2023-2026)” (arXiv, 15 July 2026; single-author preprint): https://arxiv.org/html/2607.14035v1
• SEO Vendor (Elisa Murphy), “ChatGPT Citation Gap: New Study Finds API Only Partially Matches” (21 October 2026; secondary summary): https://seovendor.co/chatgpt-citation-gap-new-study-finds-api-only-partially-matches/
• Google Search Central, “AI features and your website”: https://developers.google.com/search/docs/appearance/ai-features
• Google Search Central, “Guidelines for sites with explicit content”: https://developers.google.com/search/docs/specialty/explicit/guidelines
• Google Search Central, “Verifying Googlebot and other Google crawlers”: https://developers.google.com/search/docs/crawling-indexing/verifying-googlebot
• Google Search Central, “Do you need an SEO?”: https://developers.google.com/search/docs/beginner/do-i-need-seo