Who ChatGPT and Perplexity actually cite
A weekly measurement of which websites answer engines use as sources for this market's questions, read every Monday from a fixed published basket. This is the week of 2026-10-12. Our own domain is included on the same terms as everyone else's.
This week's readings
48
answers measured this week
IT Master citation probes, week of 2026-10-12
176
distinct sources cited
IT Master citation probes, week of 2026-10-12
Published under CC BY 4.0. Quote it, chart it, link back.
Answer engines do not rank pages. They choose sources.
Which sources they choose is measurable, and almost nobody measures it: the SEO industry writes about generative engine optimisation while treating the engines themselves as a black box. This index opens the box a little, one fixed basket of questions asked every week, with every cited source recorded.
In one sentence
Answer engines do not rank pages, they choose sources, and this index measures which sources they choose, week by week, for a fixed basket of questions this market actually asks.
The most-cited sources across both engines, ours included, whether or not we appear.
48 answers to 24 questions across both engines, week of 2026-10-12. A domain scores when it is cited at least once; the count of answers citing it and its share of the total are both shown.
| # | Domain | Share of answers | Change on last week |
|---|---|---|---|
| 1 | semrush.com | 16.7%8 answers | +4.2 |
| 2 | ahrefs.com | 16.7%8 answers | -4.1 |
| 3 | searchengineland.com | 12.5%6 answers | +4.2 |
| 4 | hubspot.com | 10.4%5 answers | -8.4 |
| 5 | arxiv.org | 10.4%5 answers | +2.1 |
| 6 | seo.com | 10.4%5 answers | +4.2 |
| 7 | reddit.com | 8.3%4 answers | 0 |
| 8 | surferseo.com | 8.3%4 answers | +6.2 |
| 9 | ai-search-tools.com | 6.2%3 answers | 0 |
| 10 | wikipedia.org | 6.2%3 answers | 0 |
| 11 | techradar.com | 6.2%3 answers | +4.1 |
| 12 | rank.ai(new this week) | 6.2%3 answers | +6.2 |
| 13 | searchable.com | 6.2%3 answers | +2 |
| 14 | chrome.com | 4.2%2 answers | +2.1 |
| 15 | answerline.dev | 4.2%2 answers | +2.1 |
| 16 | llmpulse.ai | 4.2%2 answers | +2.1 |
| 17 | frase.io | 4.2%2 answers | -2 |
| 18 | astroseoblog.com | 4.2%2 answers | +2.1 |
| 19 | subscribepr.com | 4.2%2 answers | +2.1 |
| 20 | prismic.io | 4.2%2 answers | +2.1 |
| 21 | seotakeoff.com(new this week) | 4.2%2 answers | +4.2 |
| 22 | google.cn | 4.2%2 answers | +2.1 |
| 23 | trakkr.ai(new this week) | 2.1%1 answer | +2.1 |
| 24 | enginely.ai(new this week) | 2.1%1 answer | +2.1 |
| 25 | frontiersin.org(new this week) | 2.1%1 answer | +2.1 |
| 26 | marqops.com(new this week) | 2.1%1 answer | +2.1 |
| 27 | aclanthology.org(new this week) | 2.1%1 answer | +2.1 |
| 28 | eesel.ai(new this week) | 2.1%1 answer | +2.1 |
| 29 | thecmo.com(new this week) | 2.1%1 answer | +2.1 |
| 30 | busyworth.com(new this week) | 2.1%1 answer | +2.1 |
| 31 | coursera.org | 2.1%1 answer | 0 |
| 32 | airops.com | 2.1%1 answer | 0 |
| 33 | gogochimp.com(new this week) | 2.1%1 answer | +2.1 |
| 34 | awstatic.com(new this week) | 2.1%1 answer | +2.1 |
| 35 | jasper.ai(new this week) | 2.1%1 answer | +2.1 |
| 36 | pikaseo.com(new this week) | 2.1%1 answer | +2.1 |
| 37 | pagerefresh.io(new this week) | 2.1%1 answer | +2.1 |
| 38 | keywordgraph.com(new this week) | 2.1%1 answer | +2.1 |
| 39 | besthunt.ai(new this week) | 2.1%1 answer | +2.1 |
| 40 | commoncrawl.org(new this week) | 2.1%1 answer | +2.1 |
ChatGPT and Perplexity barely agree, which is why each gets its own table.
Each engine runs its own search and picks its own sources, and for the same question the two often cite different publishers. A brand can be highly visible in one and absent from the other.
chatgpt
| # | Domain | Share of answers | Change on last week |
|---|---|---|---|
| 1 | arxiv.org | 20.8%5 answers | +4.1 |
| 2 | ahrefs.com | 20.8%5 answers | -4.2 |
| 3 | reddit.com | 16.7%4 answers | 0 |
| 4 | ai-search-tools.com | 12.5%3 answers | 0 |
| 5 | wikipedia.org | 12.5%3 answers | 0 |
| 6 | rank.ai(new this week) | 12.5%3 answers | +12.5 |
| 7 | answerline.dev | 8.3%2 answers | +4.1 |
| 8 | techradar.com | 8.3%2 answers | +4.1 |
| 9 | astroseoblog.com | 8.3%2 answers | +4.1 |
| 10 | subscribepr.com | 8.3%2 answers | +4.1 |
| 11 | searchengineland.com | 8.3%2 answers | 0 |
| 12 | google.cn | 8.3%2 answers | +4.1 |
| 13 | trakkr.ai(new this week) | 4.2%1 answer | +4.2 |
| 14 | frontiersin.org(new this week) | 4.2%1 answer | +4.2 |
| 15 | marqops.com(new this week) | 4.2%1 answer | +4.2 |
perplexity
| # | Domain | Share of answers | Change on last week |
|---|---|---|---|
| 1 | semrush.com | 29.2%7 answers | +12.5 |
| 2 | hubspot.com | 16.7%4 answers | -12.5 |
| 3 | searchengineland.com | 16.7%4 answers | +8.4 |
| 4 | seo.com | 16.7%4 answers | +8.4 |
| 5 | surferseo.com(new this week) | 12.5%3 answers | +12.5 |
| 6 | ahrefs.com | 12.5%3 answers | -4.2 |
| 7 | frase.io | 8.3%2 answers | 0 |
| 8 | searchable.com | 8.3%2 answers | 0 |
| 9 | enginely.ai(new this week) | 4.2%1 answer | +4.2 |
| 10 | eesel.ai(new this week) | 4.2%1 answer | +4.2 |
| 11 | thecmo.com(new this week) | 4.2%1 answer | +4.2 |
| 12 | chrome.com(new this week) | 4.2%1 answer | +4.2 |
| 13 | coursera.org | 4.2%1 answer | 0 |
| 14 | airops.com | 4.2%1 answer | 0 |
| 15 | gogochimp.com(new this week) | 4.2%1 answer | +4.2 |
Being cited is not the same as ranking. Domains appear here that do not rank on page one for the same question, and page-one domains are routinely absent. The overlap is real but partial, and a handful of publishers take a large share of all citations, which is the whole strategic problem for a newcomer, and why we publish our own position honestly rather than a flattering one.
Concentration is the real barrier to entry, not any single week's ranking.
Share of answers that cite at least one of the week's five most-cited domains, by week. Higher means a harder market to break into.
| Week | Answers | Sources | Top-5 concentration |
|---|---|---|---|
| 2026-09-28 | 48 | 180 | 45.8% |
| 2026-10-05 | 48 | 162 | 43.8% |
| 2026-10-12 | 48 | 176 | 39.6% |
The basket, stated in full, so the numbers can be checked rather than taken on trust.
Series 2. These 24 questions are asked verbatim, to both engines, every week. Changing the basket, or how an engine is asked or read, starts a new numbered series rather than splicing two.
- best AI SEO tools
- best AI SEO agency
- what is generative engine optimization
- what is answer engine optimization
- how to get cited by ChatGPT
- how to rank in Google AI Overviews
- best AI content writing tools for SEO
- SEO automation software
- how to track AI search visibility
- does Google penalise AI generated content
- best LLM SEO tools
- how to measure brand mentions in ChatGPT
- AI SEO tools for agencies
- what is llms.txt
- should I block GPTBot
- how to write content that AI assistants cite
- programmatic SEO tools
- how to build topical authority
- AI content detection accuracy
- best alternative to Surfer SEO
- how much does AI content cost
- E-E-A-T for AI generated content
- answer engine optimization services
- AI visibility tracking tools
A probe that errors is stored as an error, never as a zero.
'We could not measure' and 'nobody was cited' are different claims, and the method keeps them apart.
- The basket is fixed and published above. Changing it, or changing how an engine is asked or read, starts a new numbered series rather than splicing two. Series 2 kept the same questions: it began when Perplexity retired its Sonar API in September 2026, and a Perplexity source now counts only when the answer cites it, not whenever it appears in the source list. Series 2 also requires ChatGPT to search the web on every question. Its first run, on 28 September 2026, did not: ChatGPT answered 12 of the 24 questions from memory, and those probes were recorded as errors. That week was re-measured within the same week under the series-2 method, so every series-2 week is measured the same way; the replaced run is archived and available on request.
- Each question is asked verbatim, once per engine per week: to ChatGPT through OpenAI's API (the gpt-5-mini model) with web search required on every question, and to Perplexity through its Agent API, which always searches.
- What this does not measure: how often an assistant decides to search. In the ChatGPT app the assistant chooses, and on some definitional questions it answers from memory with no sources. The index measures who is cited when the assistant does go to the web, the answers that carry links and send visitors, and the model we query through the API is not necessarily the one the ChatGPT app serves you.
- Every source URL the answer cites is recorded, reduced to a registrable domain, and de-duplicated within one answer.
- A domain's share is the percentage of answers in which it was cited at least once.
- A probe that errors is stored as an error, never as a zero. A week in which an engine answered fewer than 90% of the questions is marked partial and shows no movement.
- Our own domain is measured identically and is never excluded or promoted.
- 28 Sep 2026, re-measured 28 Sep 2026: ChatGPT's web search made mandatory in series 2. Against the replaced run, the domains Perplexity cited for each question overlapped by 73% on average (its method did not change, so that is run-to-run noise) and ChatGPT's by 13% on the 12 questions it searched both times; on the 12 it had answered from memory it now cites 5.2 sources an answer, with hubspot.com and backlinko.com among them.
Questions people ask
How is this measured?
Every Monday we ask a fixed basket of 24 questions, published in full on this page, to ChatGPT with web search required on every question and to Perplexity, then record every source each answer cites. A domain scores when it appears at least once in an answer. Neither the basket nor the way an engine is asked or read changes without starting a new numbered series, so week-to-week movement means something.
Why 'share of answers' rather than share of links?
Because an engine citing one publisher three times in a single answer is one endorsement, not three. Counting links would reward publishers who happen to be quoted repeatedly in a single response and would make the numbers move for reasons that have nothing to do with trust.
Is IT Master's own site excluded?
No. Our domain is measured on exactly the same footing as every other and appears in the table when it is cited, which for most weeks it is not. Excluding ourselves would make the index worthless as evidence, and promoting ourselves would make it an advertisement rather than a measurement.
Why do ChatGPT and Perplexity differ so much?
Each engine runs its own search and chooses its own sources, and for the same question the two often cite different publishers, so a brand can be visible in one and absent from the other. Treating 'AI search' as one channel hides that, which is why the index reports each engine separately as well as combined.
Does ChatGPT always search the web?
No. The ChatGPT app decides for itself whether to search, and on definitional questions it sometimes answers from memory and shows no sources at all. This index requires a search on every question, so it measures whom ChatGPT cites when it does search — the answers that carry links — and asks both engines on the same footing. It does not measure how often ChatGPT chooses to search, and the model we query through the API is not necessarily the one the ChatGPT app serves you. An answer that searched and still cited no source counts, and lowers every domain's share.
Can I use this data?
Yes. It is published under CC BY 4.0. Quote it, chart it, write about it. A link back to this page is the only thing we ask, and it lets readers check the method for themselves.
Want the same measurement for your own domain and your own questions? Run the free AI Visibility Checker, or read how generative engine optimisation actually works.
On the list? Take the badge.
If an answer engine cites you for the questions this market asks, that is worth saying. The badge reads its rank from the live table; it will say “not in the top 40” if you are not, so it is worth exactly what it claims.
Preview
Paste this on your site
<a href="https://itmaster.uk/research/ai-citation-index">
<img src="https://itmaster.uk/research/ai-citation-index/badge/semrush.com"
alt="semrush.com in the IT Master AI Citation Index" height="28">
</a>Take the data
Published under CC BY 4.0. Use it, re-analyse it, disagree with it publicly, just say where it came from. The question basket is fixed and published above, so the numbers can be checked rather than taken on trust.
How to cite this
IT Master. “AI Citation Index, week of 2026-10-12.” itmaster.uk/research/ai-citation-index. CC BY 4.0.
We publish this because we can measure it
IT Master writes and publishes the pages that earn citations, then runs this probe nightly against your own queries.