On this page
- The short version
- Earlier studies
- Also found
- How I did it
- 0166% of the pages ChatGPT cited were published by agencies
- 02Agency-written lists were 29% of ChatGPT's citations, and ChatGPT recommended their authors 35% of the time
- 03On ChatGPT, Clutch was cited in 58% of the answers that recommended agencies
- 04When ChatGPT cited an agency's own page, it recommended that agency 93% of the time
- 0511% of the agencies ChatGPT recommended had their own site on Google's first page
- 0610% of ChatGPT's prompt-agency pairs held in all five runs
- What I would test
- Limits
- Data
- Cite this study
- About the author
Ask ChatGPT which SEO agency to hire and it gives you a shortlist with sources. I wanted to know where that shortlist comes from. So I asked ChatGPT and Google the same 300 agency-hiring questions, five times each over one day, and traced every page they cited and every agency they named.
The short version
66% of the pages ChatGPT cited when it recommended SEO agencies were published by agencies, and when it cited an agency's "best agencies" list, it recommended the agency that wrote the list only 35% of the time.
A snapshot shows what AI cites and recommends, not what changes it, so the points below are bets to test, not rules. Under each one are the numbers behind it and the finding they come from. The full version, with the strongest objection to each bet, is in What I would test.
- BetFor city searches, be on the directory pages ChatGPT cites.Evidence in Finding 3:
- For city searches, 83% of ChatGPT's citation links went to directories, against 14% for industry searches.
- Clutch was cited in 58% of the ChatGPT answers that recommended agencies.
- BetA "best agencies" list earns citations, not recommendations.Evidence in Finding 2:
- When ChatGPT cited an agency's list, it recommended the agency that wrote it 35% of the time. On Google AI Overviews, 33%.
- RedSEO's lists were cited in 104 ChatGPT answers, and ChatGPT recommended RedSEO in none of them.
- BetSay plainly on your homepage and industry pages what you do and for whom.Evidence in Finding 4:
- When ChatGPT cited an agency's homepage, industry, service or case study page, it recommended that agency in the same answer 93% of the time.
- The recommended agencies' own pages, not counting their lists, were 32% of what ChatGPT cited when it recommended agencies.
- BetGoogle rankings will not carry you on ChatGPT.Evidence in Finding 5:
- 11% of the agencies ChatGPT recommended had their own site on Google's first page for the same query. On Google AI Overviews, 32%.
- BetJudge AI visibility on repeated runs, never on one screenshot.Evidence in Finding 6:
- 10% of ChatGPT's prompt-agency pairs came back in all five runs.
What earlier studies found
- Glen Allsopp at Ahrefs looked at the sources ChatGPT cited for software, product and agency questions and found that "best X" blog lists were 43.8% of the page types ChatGPT cited, across 750 prompts (Ahrefs). He also found that brands placed higher on the cited lists were recommended more often.
- Lily Ray tracked Google's AI Overviews for business software questions and found that when a brand's own "best" list was cited, the brand was left out of the recommendation 69% of the time (224 of 323 cited lists) (Lily Ray's newsletter).
- Mateusz Makosiewicz at Ahrefs published self-promotional pages as an experiment and found that when AI cited a page promoting Ahrefs' own conference, it still skipped the conference 43% of the time (Ahrefs).
- Louise Linehan at Ahrefs compared AI citations with Google's results for the same prompts and found that on average only 12% of the links ChatGPT, Gemini and Copilot cited were in Google's top 10 for the same prompt, across 15,000 prompts (Ahrefs).
What's new here: Lily Ray measured how often a list's author is left out on Google, for software lists that rank their own brand. This study asks it of every agency-written list for SEO agencies, on ChatGPT as well as on Google's AI Overviews, and across five runs of every question: the author was left out 67% of the time on AI Overviews and 65% on ChatGPT. It also traces who wrote everything else the answers cited, which directory owns city searches, and how much of each shortlist survives a rerun.
Also found
- Next to a recommended agency, ChatGPT most often linked the agency's own page. Own page 34%, listicle 30%, directory 20% (Finding 4).
- The agency's own site was missing from the sources of 59% of ChatGPT's recommendations. Every visible source for those was a third-party page (Finding 4).
- Google's AI Overviews leaned on one agency's lists. Thrive Internet Marketing Agency was cited in 28% of AI Overviews, mostly its "best X SEO companies" lists and its city pages (Finding 3).
How I did it
- The questions 300 SEO agency-hiring questions I wrote, for city, industry, specialty and buyer-need searches, plus plain service searches that are left out of the findings (Data).
- The surfaces ChatGPT and Google's AI Overviews, every question asked five times on 13 September 2026, roughly an hour apart.
- What I recorded Every page each answer cited and every agency it recommended, then who published each cited page, what kind of page it was, and whether the recommended agency wrote it.
Finding 0166% of the pages ChatGPT cited were published by agencies
When ChatGPT recommended SEO agencies, here is who published the pages it cited:
| Publisher of the cited page | ChatGPT | Google AI Overviews |
|---|---|---|
| The recommended agency's own site | 42% | 28% |
| Another agency's site | 24% | 40% |
| Directory or review marketplace (Clutch, Expertise, DesignRush) | 24% | 6% |
| Publisher, media, software vendor | 7% | 4% |
| Google Business Profile card | 0% | 10% |
| Reddit, YouTube, LinkedIn | 0% | 9% |
| Could not be classified | 3% | 3% |
Combined, the first two rows are 66% of ChatGPT's citations and 68% of Google's coming from agency websites. Each row is rounded on its own, so the rounded rows may not add up exactly. The 95% uncertainty range for the ChatGPT figure is 63% to 69%. Independent sources are a small minority. On ChatGPT, publishers, media and software vendors account for 7% of citations. On Google, Reddit and YouTube together account for 6%, and all community sites including LinkedIn for 9%.
This matters because when an AI answer says "according to these sources," the sources are mostly agencies describing themselves and each other.
Finding 02Agency-written lists were 29% of ChatGPT's citations, and ChatGPT recommended their authors 35% of the time
The single most cited page type on both surfaces is the list article: "The 10 Best Technical SEO Agencies in 2026," "Best SEO Companies for Small Business," and so on.
| Page format | ChatGPT | Google AI Overviews |
|---|---|---|
| "Best agencies" list or comparison article | 38% | 41% |
| Directory category page | 24% | 6% |
| Homepage | 17% | 3% |
| Industry landing page | 6% | 9% |
| Service page | 4% | 4% |
| Case study | 4% | 0% |
| How-to or guide article | 2% | 9% |
| City landing page | 1% | 6% |
Most of those lists are written by agencies. On ChatGPT, 1,891 citation links (29% of all citations) went to list articles on agency websites. I then checked whether the agency that wrote the list was itself recommended in the same answer.
On ChatGPT it was, 35% of the time. In the other 65% of cases, ChatGPT cited an agency's list but did not recommend the agency that wrote it. On Google AI Overviews, agency-written lists are 36% of citations and the author is recommended 33% of the time.
It depends a lot on who you are. Each bar is one agency's lists: how often ChatGPT also recommended the agency that wrote them, with the number of ChatGPT answers that cited them.
RedSEO's comparison pages were cited in 104 ChatGPT answers. ChatGPT recommended RedSEO in none of them. The same pattern shows up on Google: Web Tonic's lists were cited in 91 AI Overviews and Web Tonic was recommended in 10.
I checked the cited list pages. On the lists I checked, Searchbloom, Thrive, First Page Sage, Siege Media and SEO.com all put themselves on their own lists, usually in first place. But listing yourself first is not enough: RedSEO ranks itself #1 on all seven of its cited lists, and ChatGPT recommended RedSEO in none of the 104 answers that cited them. SEO Engico ranks itself #1 on all 11 of the lists ChatGPT cited and was recommended in 1 of 35 of those ChatGPT answers. At the other extreme, PikaSEO leaves itself off all three of the lists ChatGPT cited, and The Rank Masters leaves itself off four of its five ChatGPT-cited lists.
I did not find what does explain the difference, and this study cannot: it is a snapshot, not an experiment. What it does show is that a cited list is not the same as a recommendation.
(One note on the SEO.com row: SEO.com describes itself as "powered by the experts at WebFX," so I counted a WebFX recommendation as a self-recommendation for SEO.com's lists.)
Finding 03On ChatGPT, Clutch was cited in 58% of the answers that recommended agencies
Clutch appeared in the sources of 569 of the 988 ChatGPT answers that recommended agencies (58%). Expertise.com appeared in 25% of those ChatGPT answers. No agency website came close: the most cited agency domain on ChatGPT, SEO.com, appeared in 20%.
Directory dependence is almost entirely a city phenomenon:
| Buyer context (ChatGPT) | Directory share of citation links | Agency-site share | Top domains |
|---|---|---|---|
| City searches | 83% | 15% | clutch.co, expertise.com, seo.com, designrush.com, seos.co |
| Industry searches | 14% | 75% | firstpagesage.com, clutch.co, seos.co, thestacc.com |
| Specialty searches | 17% | 69% | clutch.co, searchbloom.com, thestacc.com, victorious.com |
| Buyer-need searches | 12% | 77% | clutch.co, redseo.io, thestacc.com, seo.com |
Ask ChatGPT for the best SEO agencies in Nashville and it cites Clutch's Nashville page and Expertise's Nashville page, in all 10 answers I collected. Ask ChatGPT for the best SEO agencies for dental practices and 74% of its citations are agency-written lists instead.
Google AI Overviews behave differently. In AI Overviews, directories are only 6% of citations overall and 9% even for city searches. Google leans on Thrive Internet Marketing Agency (cited in 28% of AI Overviews, mostly its "best X SEO companies" lists and its city pages), Reddit threads (19%), YouTube (19%) and its own Business Profile cards (10% of citation links).
Finding 04When ChatGPT cited an agency's own page, it recommended that agency 93% of the time
In answers that recommended agencies, the recommended agencies' own pages were 42% of everything ChatGPT cited, or 32% not counting their own lists. Of those non-list pages ChatGPT cited, homepages were 50%, industry landing pages 18%, service pages 12% and case studies 12%.
Own pages show up mostly outside city searches. Not counting lists, they were 33% of what ChatGPT cited for industry searches, 36% for specialty searches, 41% for buyer-need searches and 4% for city searches. On Google AI Overviews they were 16% of citations, mostly industry and city pages. City landing pages were almost invisible on ChatGPT (1% of all citations) but mattered on Google (6%), and Thrive's city pages alone were 14% of those.
When ChatGPT cited an agency's own homepage, industry page, service page or case study, that same agency was recommended in the same answer 93% of the time. When ChatGPT cited an agency's "best agencies" list, the list's author was recommended 35% of the time. Google AI Overviews show the same pattern, less strongly:
| Agency page that was cited | Same agency recommended, ChatGPT | Same agency recommended, Google AI Overviews |
|---|---|---|
| Homepage | 96% | 73% |
| Industry landing page | 91% | 56% |
| Service page | 90% | 63% |
| Case study | 90% | Not cited |
| "Best agencies" list | 35% | 33% |
These numbers show that citations and recommendations appear together, not which comes first: an answer may recommend an agency and then link its site. Finding out whether improving a page changes who gets recommended needs a before-and-after test.
But an agency can be recommended without any page of its own being cited. Across 5,312 ChatGPT recommendations I could match to a website, the agency's own site was in the source list 41% of the time. On Google AI Overviews, 42%. For the other recommendations, every source shown was a third-party page, such as a directory, another agency's list or a community thread.
These ChatGPT figures count different things: 41% of recommendations had the agency's own site cited, while 32% of citation links were the recommended agencies' own pages, not counting lists.
ChatGPT also usually puts a source link right next to each agency it recommends. Across 5,312 ChatGPT recommendations and 2,974 AI Overview recommendations I could match to a website, this is what was linked next to the agency's name. Each recommendation counts once. When several kinds of page were linked, the agency's own page counts first, then a listicle, then a directory.
| Linked next to the recommended agency | ChatGPT | Google AI Overviews |
|---|---|---|
| The agency's own page | 34% | 21% |
| A listicle, the agency's own or another site's | 30% | 35% |
| A directory | 20% | 4% |
| Some other page | 1% | 5% |
| Nothing linked next to the agency | 15% | 35% |
On ChatGPT, own pages sit next to more recommendations than listicles do, and in 81% of the listicle cases the listicle belonged to another site. On Google AI Overviews, listicles sit next to more recommendations than own pages do. "Nothing" means no link sat next to the agency's name. The answer may still cite sources elsewhere. This measure is narrower than the 41% of recommendations above for ChatGPT, which counts the agency's site anywhere in the answer's sources, including its own listicles.
Finding 0511% of the agencies ChatGPT recommended had their own site on Google's first page
For every ChatGPT answer I also had Google's first page of results for the same query, collected at the same time.
- 8% of the URLs ChatGPT cited were on Google's first page for that query.
- 15% of the domains ChatGPT cited were on that first page.
- 11% of the agencies ChatGPT recommended had their own site on that first page.
For Google's own AI Overviews the overlap is higher, as you would expect: 50% of cited domains were on the first page, and so were 32% of recommended agencies' own sites.
Louise Linehan at Ahrefs measured the same overlap across AI assistants and long-tail questions: on average only 12% of the links ChatGPT, Gemini and Copilot cited were in Google's top 10 for the same prompt, across 15,000 prompts. For these agency questions, 8% of the URLs ChatGPT cited were on Google's first page.
If your AI visibility strategy is "rank well and the rest follows," note that ranking and being recommended go together partly on Google and barely on ChatGPT.
Finding 0610% of ChatGPT's prompt-agency pairs held in all five runs
I ran every prompt five times, roughly an hour apart. On ChatGPT, the median overlap of recommended agencies between one run and the next was 0.30 (Jaccard index: shared agencies divided by all agencies named in either run). For cited domains on ChatGPT it was 0.33.
On ChatGPT, out of 2,754 prompt-agency pairs, 263 (10%) appeared in all five runs, and 1,381 (50%) appeared once.
Google AI Overviews were steadier: 0.43 for agencies, 0.43 for domains. Among the 140 questions that showed an AI Overview in all five runs, 198 of 1,091 prompt-agency pairs (18%) held for all five.
Which agencies were consistent? On ChatGPT, WebFX was recommended in 26% of the ChatGPT answers that recommended agencies, Victorious in 24%, Searchbloom in 22%, SmartSites in 17%, Siege Media in 15%. Each of those five was recommended across a wide spread of cities, industries and specialties. Below the top ten on ChatGPT the distribution is very flat: it takes 39 agencies to cover half of all ChatGPT recommendations, and 923 distinct agencies were recommended outside the local business cards.
If you are reporting AI visibility to a client from one screenshot, most of what you see can change on the next run.
What I would test
A snapshot shows what AI cites and recommends, not what changes it, so each point below is something to test rather than a rule.
Check the exact pages AI reads for your market, then test whether getting onto them changes anything. For city searches on ChatGPT, that is your Clutch and Expertise category pages. Those are the pages to check your presence on. For industry and specialty searches on ChatGPT, agency-written lists make up 38% and 30% of citation links, spread across many sites. Find the ones cited for your niche, and find out whether you are on them.
If you publish a "best agencies" list, test what it does for you, not just whether it gets cited. RedSEO ranks itself #1 on all seven of the lists ChatGPT cited and was never recommended when those lists were cited. A citation is not a recommendation. Owned lists do go with more recommendations, though: in runs of the same question, ChatGPT recommended an agency 28% of the time when its own list was cited and 10% when it was not. Agencies ChatGPT also recommended in answers without their list were recommended in 55% of the answers citing their own list, the rest in 8%.
Test whether plain words on your homepage and industry pages change what AI says about you. On ChatGPT, after lists and directories, they are the pages cited most. When ChatGPT cites an agency's homepage, industry, service or case study page, it recommends that agency in the same answer 93% of the time. That shows the pages and the recommendation go together, not which comes first. "Full-service digital growth partner" does not help a model decide whether you do dental SEO in Ohio.
Test before you assume Google rankings carry over. In this data they partly do for AI Overviews and barely do for ChatGPT.
Test with repeated runs, not one screenshot. Five runs an hour apart gave materially different shortlists. Any AI visibility report should say how many runs it is based on.
What this study does not show
The prompts were written by me, not pulled from real user logs. The 300 prompts are a deliberate commercial sample, not a ranking of what buyers actually type.
Agency names were extracted from answers with rules, not by hand. A hand-labelled check of 100 randomly drawn answers put the rules' precision at 98% and their recall, the share of offered businesses they caught, at 92%. The hand labels in that check and in the publisher-label check below were made by Claude, an AI assistant, without sight of the pipeline's output. Matching each agency to its website was automated too. An agency I could not match could not be linked to citations, and a wrong match links the wrong site, so the own-site and self-recommendation figures can be off in either direction. Checks on 16 to 18 September 2026 corrected 1,544 of the 21,444 extracted answer-agency pairs, including removing headings, directories, software and articles that had been extracted as if they were agencies, and separating companies the extractor had merged. For 1,186 of the removed pairs, an AI model flagged the name as not an agency; in a random sample checked by hand, its label was right for 62 of 65 names. Other matching errors may remain.
Links next to each agency were assigned by rules, from the list items, table rows or sections that start with the agency's name, or from the line that names it when there is no such item. A hand check of 20 randomly chosen recommendations found no misassigned links.
Publisher labels were reviewed domain by domain for the domains that make up 80% of citations. The remaining long tail was labelled automatically, from page titles or from matching a recommended agency's website. A hand check of that tail found 89 of 100 randomly drawn pages labelled correctly, and the wrong ones were corrected. In the reviewed group, title-based labels were right 77% of the time, so if the title-based labels in the tail are just as accurate, the true agency-published share is about 1 point lower than reported for ChatGPT and about 3 points lower for AI Overviews. Page-format labels were checked on a stratified sample of 110 pages, with 104 correct.
Nothing here is causal. I measured what two AI surfaces cite and name for a fixed panel of questions on one day. Whether getting on a Clutch page or writing a list changes what they say is a question for an experiment, not a snapshot.
One artifact worth knowing about: both ChatGPT and Google sometimes open an answer with Google-Maps-style business cards. For prompts with no city in them, ChatGPT's cards showed agencies in locations unrelated to the buyer. I flagged those and left them out of the headline agency figures.
Data
3,000 searches: 300 questions, run five times each on both platforms. Every finding in this post comes from the questions that ask for the best agencies or whom to hire: 1,000 ChatGPT answers and 866 AI Overviews. The other 100 questions were plain service searches, such as "SEO agencies in New York, NY", and are left out of the findings, since an answer to a general service search is not necessarily a recommendation.
A citation link is a page in an answer's source list, counted once per answer. A recommendation is an agency the answer names as an option, counted once per answer.
In the 988 ChatGPT answers that recommended agencies: 6,440 citation links across 850 domains. In the 830 AI Overviews that recommended agencies: 6,150 citation links across 789 domains. Across all answers, including the plain service searches and local business cards, there were 20,200 answer-agency pairs. Every percentage in this post has a numerator and a denominator in the study's metrics file, and every answer is preserved as the raw provider response.
Study by Ivan Smiljkovic. Data collected on 13 September 2026. Every figure has a numerator and a denominator in the study dataset; every raw answer is preserved.
Cite this study
Ivan Smiljkovic (2026). 66% of the Pages ChatGPT Cites When Recommending SEO Agencies Are Written by SEO Agencies (Study of 3,000 AI Searches). Ivan Builds. https://ivanbuilds.com/ai-agency-citations.html
Every section has its own link: hover a heading and use the # to copy it.
About the author
I study how Google and AI assistants choose what to cite, using public data I collect myself, and I publish the method and limits with every study.