AI answers from ChatGPT, Claude, Gemini, Perplexity and Google AI Overviews
Ask ChatGPT for software and its web search does more than repeat your question. It rewrites it, often adding vendor names you never mentioned. Every Monday, Fast Growth Advisors puts the same 40 B2B buying queries, from HR (human resources) to cybersecurity, to ChatGPT, Claude, Gemini, Perplexity and Google AI Overviews, on the British market. Our AI answers barometer compares what each engine searches for, what it cites and what holds from one week to the next.
What do ChatGPT, Claude, Gemini, Perplexity and Google answer this week?
Updated on 23 Sep 2026. Latest week published: 21 Sep 2026 to 27 Sep 2026. Series started on 23 Sep 2026. A week is published once its run has finished.
The series is starting: every Monday adds a point. Charts appear from the second week, and source stability from the second measurement of the same query.
Three tables: the assistants against each other, their sources against Google’s, then Google alone.
- ChatGPT
- Claude
- Gemini
- Perplexity
- Google AI Overviews
The four assistants
Answers that trigger a web search
In grey: Answers given from memory, without a web search
Sources cited per answer that searched
Searches phrased in another language than the query
Query words kept in the search
Searches that add a brand absent from the query
Sources kept from one week to the next, same query
no data
| Indicator | ChatGPT | Claude | Gemini | Perplexity |
|---|---|---|---|---|
| Answers that trigger a web search | 60% | 100% | 80% | 100% |
| Answers given from memory, without a web search | 40% | 0% | 20% | 0% |
| Sources cited per answer that searched | 5.9 | 14.4 | 9.6 | 17.4 |
| Searches phrased in another language than the query | 0% | 0% | 0% | not exposed |
| Query words kept in the search | 57% | 86% | 82% | not exposed |
| Searches that add a brand absent from the query | 51% | 0% | 4% | not exposed |
| Sources kept from one week to the next, same query | no data | no data | no data | no data |
Do the assistants cite the same sources as Google?
Share of their sources that Google AI Overviews also cites, same query, same week
| Indicator | ChatGPT | Claude | Gemini | Perplexity |
|---|---|---|---|---|
| Share of their sources that Google AI Overviews also cites, same query, same week | 12% | 16% | 22% | 17% |
Google AI Overviews
| Indicator | Google AI Overviews |
|---|---|
| Queries where Google shows an AI Overview | 100% |
| Sources cited per AI Overview shown | 7.7 |
| Sources kept from one week to the next, same query | no data |
What the engines actually searched for
Query asked: best hr software for small business
ChatGPT: best HR software for small business 2026 BambooHR Personio CharlieHR Breathe Sage HR Zoho People Rippling comparison
Brands added by ChatGPT: bamboohr, charliehr, personio, rippling, sage, zoho
Claude: best HR software for small business 2026
Gemini: best hr software for small business
Query asked: ai enterprise search software
ChatGPT: enterprise AI search software vendors 2026 Coveo Algolia Lucidworks Amazon Kendra Elastic OpenSearch Sinequa Microsoft Google Cloud Search
Brands added by ChatGPT: amazon, coveo, elastic, microsoft
Claude: best AI enterprise search software 2026
Gemini: top AI enterprise search software 2025 2026
Query asked: best endpoint protection for business
ChatGPT: EDR EPP comparison 2026 CrowdStrike Microsoft Defender for Endpoint SentinelOne Trend Micro Sophos 2026 review
Brands added by ChatGPT: crowdstrike, microsoft, sentinelone, sophos
Claude: best endpoint protection for business 2026
Gemini: best endpoint protection for business 2025 2026
Query asked: best product analytics tools
ChatGPT: best product analytics tools 2026 list Mixpanel Amplitude Heap Pendo FullStory Contentsquare
Brands added by ChatGPT: amplitude, contentsquare, pendo
Claude: best product analytics tools 2026
Gemini: top product analytics tools 2025 2026
One real query per sector, this week: the query asked, the search ChatGPT issued when it rewrote it the most, and the first search by Claude and Gemini on the same query. Examples, not a ranking of the brands named.
What does ChatGPT actually search for when asked for software?
From the barometer’s first run, on 23 September 2026, our measurements showed a clear gap between the engines. ChatGPT issues several searches for a single question, rewrites its wording far more than Claude and Gemini do, and adds vendor names the question never mentioned.
Claude and Gemini mostly keep the words of the question.
A search that already names vendors is no longer used to discover them: it checks the ones it names. For a brand, the consequence is direct. Missing from the list the engine starts searching with, it has little chance of showing up in what it finds.
Two indicators in the four assistants table track this every week: “Query words kept in the search” and “Searches that add a brand absent from the query”. The real examples under “What the engines actually searched for”, just above, change every Monday. They are neither a ranking nor a recommendation.
Do Google and the assistants cite the same sources?
On the same 40 queries, Fast Growth Advisors also records every week what Google writes in its AI Overviews and whom it cites. It is a comparison we have not seen published anywhere.
Nothing says they read the same web.
A brand well cited by Google can be missing from an assistant’s sources, and the other way round. The table “Do the assistants cite the same sources as Google?” measures it: for each assistant, the share of its sources that Google AI Overviews also cites, same query, same week. Google does not expose the searches it runs, so its own table keeps to three indicators.
Which domains do the engines cite, sector by sector?
Behind the indicators sit domain names. For each of the four sectors, from payroll to data, Fast Growth Advisors ranks the domains cited by the most answers that searched the web, all engines together, then the top five of each engine. These rankings describe what the engines cite. No judgement on the sites cited is implied.
Most cited domains, by sector
| Rank | Domain | Share of answers |
|---|---|---|
| 1 | peoplemanagingpeople.com | 28% |
| 2 | personio.com | 26% |
| 3 | zapier.com | 26% |
| 4 | hibob.com | 26% |
| 5 | youtube.com | 26% |
| 6 | techradar.com | 21% |
| 7 | peoplehr.com | 19% |
| 8 | rippling.com | 17% |
| 9 | workforce.com | 17% |
| 10 | capterra.co.uk | 15% |
| Rank | Domain | Share of answers |
|---|---|---|
| 1 | youtube.com | 32% |
| 2 | reddit.com | 25% |
| 3 | zapier.com | 25% |
| 4 | slite.com | 20% |
| 5 | hebbia.com | 14% |
| 6 | slack.com | 14% |
| 7 | cloud.google.com | 11% |
| 8 | learn.g2.com | 11% |
| 9 | eesel.ai | 11% |
| 10 | zendesk.com | 11% |
| Rank | Domain | Share of answers |
|---|---|---|
| 1 | gartner.com | 43% |
| 2 | expertinsights.com | 27% |
| 3 | pcmag.com | 25% |
| 4 | sentinelone.com | 20% |
| 5 | huntress.com | 18% |
| 6 | youtube.com | 18% |
| 7 | techradar.com | 16% |
| 8 | paloaltonetworks.com | 16% |
| 9 | manageengine.com | 14% |
| 10 | techrepublic.com | 14% |
| Rank | Domain | Share of answers |
|---|---|---|
| 1 | domo.com | 32% |
| 2 | gartner.com | 25% |
| 3 | reddit.com | 22% |
| 4 | g2.com | 22% |
| 5 | youtube.com | 20% |
| 6 | ovaledge.com | 18% |
| 7 | atlan.com | 18% |
| 8 | techtarget.com | 18% |
| 9 | solutionsreview.com | 18% |
| 10 | thoughtspot.com | 15% |
Share of answers that searched the web and cite the domain at least once, over the last measured week (175 answers).
Top five domains cited by each engine
ChatGPT
- techradar.com 42%
- gartner.com 17%
- personio.com 12%
- reddit.com 12%
- ico.org.uk 12%
Share of citations captured by the top 10 domains: 24%
Claude
- gartner.com 28%
- peoplemanagingpeople.com 22%
- capterra.com 15%
- g2.com 15%
- expertinsights.com 15%
Share of citations captured by the top 10 domains: 11%
Gemini
- youtube.com 47%
- reddit.com 31%
- zapier.com 12%
- gartner.com 12%
- saascompared.io 9%
Share of citations captured by the top 10 domains: 16%
Perplexity
- learn.g2.com 30%
- zapier.com 28%
- techrepublic.com 22%
- gartner.com 22%
- techtarget.com 22%
Share of citations captured by the top 10 domains: 12%
Google AI Overviews
- youtube.com 67%
- reddit.com 38%
- gartner.com 26%
- zapier.com 18%
- ibm.com 10%
Share of citations captured by the top 10 domains: 26%
Most cited domains and citation concentration, by engine
How are the AI answers collected?
Every Monday at 06:17 UTC (Coordinated Universal Time), Fast Growth Advisors puts the 40 queries to the four assistants through their API (application programming interface), with one fixed model each, and records Google AI Overviews through a probe located in the United Kingdom. Results are published the same day, around 10:00 UTC.
The models queried are gpt-5-mini for ChatGPT, claude-sonnet-5 for Claude, gemini-3.6-flash for Gemini and sonar for Perplexity. Consumer apps work differently: settings, memory, personalisation. A signed-in user may therefore get another answer than the one recorded here. What the barometer measures is the bare model, under fixed and documented conditions.
For ChatGPT, Claude and Perplexity, the search country is set to the United Kingdom; Gemini accepts no country setting and searches from France, including for the British queries.
Google AI Overviews is recorded around 06:20 UTC by a DataForSEO probe, and unlike Gemini through the API, its British location is real.
Written like real searches, in lower case and without punctuation, the 40 queries cover four sectors (human resources and payroll, artificial intelligence, cybersecurity, data and analytics), each with five comparison queries and five need queries. They are the British queries of the four fixed sectors of the Google AI Overviews barometer, frozen on 2 September 2026 and written for this market rather than translated from the French set.
The series started on 23 September 2026, with no history before it. The copy on this page quotes no result: a sentence repeating the figures of the tables would be wrong by the following Monday. Finally, Fast Growth Advisors’ own domains, fast-growth.fr and nomo-ia.com, are excluded from every ranking.
Google AI Overviews barometer, September 2026 edition →
What the Google AI Overviews barometer measures →
The 40 queries measured
HR and payroll
- best hr software for small business
- comparison hr software uk
- best applicant tracking system uk
- best payroll software uk
- best employee onboarding software
- hr software for growing companies
- employee time tracking software
- absence management software uk
- workforce scheduling software
- hr analytics software
Artificial intelligence
- best ai assistant for business
- comparison ai chatbot for business
- best ai coding assistant
- best ai content generation tools
- comparison ai agents for business
- ai document analysis software
- ai knowledge base software
- ai invoice processing software
- ai customer relationship tool
- ai enterprise search software
Cybersecurity
- best endpoint protection for business
- best antivirus for business uk
- comparison siem solutions for business
- best password manager for business
- best threat detection software
- vulnerability management saas
- phishing awareness training for staff
- identity and access management software
- cloud backup for business uk
- email security solution for business
Data and analytics
- best business intelligence software
- comparison etl tools
- best data visualisation tools
- best product analytics tools
- google analytics alternative gdpr
- data quality software for business
- data governance software
- data catalogue software
- data warehouse for small business
- kpi dashboard software
How are the eleven indicators calculated?
Eleven indicators, in three groups and in the order of the tables: seven for the four assistants, one comparing their sources with Google’s, three for Google AI Overviews. Fast Growth Advisors computes them engine by engine, with the same method every week.
For the four assistants.
“Answers that trigger a web search”: the share of answers for which the engine looked things up on the web rather than answering from memory.
“Answers given from memory, without a web search”: its complement, the two add up to 100%.
“Sources cited per answer that searched”: the average count of distinct domains cited.
“Searches phrased in another language than the query”: searches written in French on the British market. Those whose language cannot be determined are set aside.
“Query words kept in the search”: for each search, the share of the query’s meaningful words it contains, on average. Stop words and “best” are left out, accents ignored.
“Searches that add a brand absent from the query”.
“Sources kept from one week to the next, same query”: the Jaccard index, meaning the domains shared by two readings divided by all the domains cited in either. It equals 1 when the sources are identical and 0 when none is carried over. Two consecutive readings are compared, no more than eight days apart.
Perplexity does not publish its searches, so its column reads “not exposed” for the three indicators about them.
Between the assistants and Google. “Share of their sources that Google AI Overviews also cites, same query, same week”: for each assistant, the share of the domains it cites that Google also cites in its AI Overview.
For Google AI Overviews. “Queries where Google shows an AI Overview”, “Sources cited per AI Overview shown”, the average number of domains an AI Overview cites, and “Sources kept from one week to the next, same query”, computed as for the assistants.
Three caveats frame the brand indicator and the examples. A brand is only recognised if a domain bearing its name was cited by one of the five engines the same week, on the same market, and it is only counted if the engine wrote it as a proper noun, starting with a capital letter or a digit. Generic platforms (AI engines, YouTube, Reddit, GitHub) never count. Behaviour depends on the model: another OpenAI model, measured on 3 September 2026, slipped the site: operator into most of its searches. Finally, the examples name vendors because the engine wrote them.
Any share computed on fewer than ten items is not shown. “No data” marks a measurement that does not exist yet. Only valid host names count as domains. Each week is published only once its run has finished, and a failed run shows up as a gap, dated, with its reason.
FAQ
Why publish a weekly series rather than a ranking?
Because a generative AI answer changes from one run to the next. A ranking built on a single reading mixes trend and chance, with no way to tell them apart. When the same queries go to the same models every Monday, the gap between two weeks becomes readable, and a one-off anomaly stops looking like a lasting shift.
Why does ChatGPT add brands to my question?
The barometer measures what ChatGPT does, not what it wants. When it searches the web from a category question, a good share of its searches already name vendors that were absent from the question, whereas those of Claude and Gemini almost never do. Where those names come from, the measurement does not say. It does show the consequence for a vendor: if it does not appear in those searches, it has little chance of being found by them.
Does being cited by Google AI Overviews mean being cited by ChatGPT?
Not necessarily. Google and the assistants do not necessarily draw on the same sources: a domain cited by AI Overviews can be missing from ChatGPT’s answers, and the other way round. The barometer measures it every week: for each assistant, the share of its sources that Google AI Overviews also cites on the same query. That is the indicator to watch before concluding that visibility won in Google carries over to the assistants.
How does it differ from the Google AI Overviews barometer?
Same questions, two uses. Once a month, a dedicated edition of the Google AI Overviews barometer analyses in detail what Google writes above its results and whom it cites. Once a week, the AI answers barometer records Google and four assistants side by side, to compare them. Measurements are stored apart, so a monthly edition stays reproducible and the figures of the two pages never mix.
How should a cell without a figure be read?
Any share computed on fewer than ten items says more about chance than about the engine, so its cell shows no figure. “Not exposed” marks an engine that does not publish its searches, such as Perplexity. “No data” marks a measurement that does not exist yet, for example source carry-over while the series has only one week.
Next in this series:
Measure AI visibility, the method and the instruments →
What do the engines answer in your category?
The free diagnostic puts five questions from your category to ChatGPT, Perplexity and Google AI Overviews. Answers archived, verification link included, causes read on both the message and the pages.