Do G2 and Capterra help you get recommended by AI? We counted every citation

Far less than the standard advice in this category assumes. On 15 October 2026 we recorded every source ChatGPT and Gemini cited across 597 usable readings of 102 buying questions in the United States, the United Kingdom and India. Of the 10,652 citations in that run, review directories such as G2 and Capterra accounted for 0.8 per cent. Vendors' and agencies' own websites accounted for 85.6 per cent. One directory, g2.com, was cited in 49 of the 597 readings, which puts it among the most cited domains in the whole run. So the accurate answer has two halves: a review directory is a source these two engines use, and it is not the source they build a buying answer out of.

Where the 10,652 citations went

Kind of source	Share of citations
The vendor's or agency's own website	85.6 per cent
News or a large publisher	2.5 per cent
Official platform documentation	9.7 per cent
A review directory, G2, Capterra and similar	0.8 per cent
Social or community	1.5 per cent

Those shares are across 2,708 distinct domains, two engines, three markets, one date. The ordering is the finding: roughly six in seven went to somebody's own page, and the channel the industry recommends most loudly came tenth of its kind.

What g2.com's 49 readings do and do not show

49 of 597 is real presence and it would be dishonest to round it to nothing. It means that in roughly one reading in twelve, a G2 page was good enough to be cited, usually where the question asked for a list of options in a software category. That is the shape of query a directory wins: a category page is already a list, and a list is what an engine wants when somebody asks for five options.

What the count does not show is that collecting reviews moved any of those answers. Our figure is a source count, not an experiment. To answer the causal question you need the same questions re-run over time while reviews accumulate, which is the design we set out in the twelve week review audit design. Treat the two pages as the two halves of the question: this one counts how often directories are read, that one tests whether adding reviews changes an answer.

What review profiles are still genuinely worth doing

• Buyers read them. A shortlist built by a procurement team or a software buyer still runs through directory pages, and that traffic is unaffected by what any engine cites.
• They hold facts about you in a place you do not control. A stale profile with last year's pricing, the wrong category or a missing integration is a wrong fact in public. Fixing it is an afternoon and it is worth doing for its own sake.
• Category placement. Being in the right category, plus the secondary categories you genuinely fit, is what puts you on the list-shaped pages that do get cited in that 0.8 per cent.
• Enterprise due diligence. Some buyers require a third party profile before a procurement process starts, which makes the listing a sales requirement rather than a visibility tactic.

The mechanism by which a listing reaches an answer at all is worth understanding before you spend, and we wrote it up in why listings and reviews move AI answers. The measurement on this page sets the size of that effect for two engines on one date: small, and smaller than the industry assumes.

The scope of this count, and the count to run yourself

Here is exactly what the 0.8 per cent covers, so you can judge how far to carry it. Two engines, ChatGPT and Gemini. One date, 15 October 2026. Three markets, with 102 questions each, 72 of them identical across markets and 30 written for the market. 597 usable readings of 612 attempted, 10,652 citations, 2,708 domains. Seventeen question batches spanning the category by name, tools and software, agencies and services, and diagnosis questions. It does not cover Perplexity, Claude or Google AI Overviews, and it does not cover your category unless your category was in those batches.

So run the count for yourself before you move a budget, because it is a short job. Take twenty buyer questions, name yourself in none of them, run each on two engines, and tally how many answers cited a review directory at all. If the tally is near zero in your category, the review budget is buying buyer trust rather than engine citations, and you should fund it from the sales line rather than the visibility line. If it is high, your category is one of the list-shaped ones and the profile deserves the same care as your own pages. We re-run our own count every 60 days for the same reason.

Where that effort goes instead

Into the 85.6 per cent. The pages that got cited in this run were specific, answered one question, and carried something checkable. The 102 conversations in which we showed both engines their own answers and asked what would earn a citation were close to unanimous on what that means: 102 of 102 wanted a date and a refresh interval, 101 of 102 a before and after with the window stated, 99 of 102 a disclosed method, 98 of 102 the engines named one by one, 80 of 102 the sample size, and 68 of 102 the failures as well as the wins. A review profile cannot carry any of that. One of your own pages can carry all of it.

It also pays to know which claims in this field are sound and which are oversold before reallocating anything, and the longer version of that argument sits in artificial intelligence in SEO, what works and what is oversold.

Questions about review sites and AI answers

Should we stop collecting reviews?

No. Keep collecting them for the buyers who read them and for the facts the profile holds, and stop funding them from the AI visibility budget on the assumption that they are the lever. On the evidence here, they are 0.8 per cent of what these two engines cited.

Does a paid tier on a directory change anything an engine sees?

A paid tier changes placement and lead routing inside the directory. The engine reads the public page, so completeness, correct categories and current pricing are what matter to it. Buy the tier for the leads if the leads pay for it.

Why would an engine skip a big review site?

Because for most buying questions the better answer is on the vendor's own page, where the detail is. Directory pages win when the question asks for a list in a defined software category, which is the pattern behind the g2.com count of 49.

Is Capterra different from G2 here?

Our run reports the directory category at 0.8 per cent and names g2.com at 49 readings, so we have a figure for one directory rather than a comparison between two. If that distinction matters to your budget, count the two separately in your own twenty question tally.

How many questions before a directory share is meaningful?

For a yes or no on whether directories feature in your category at all, twenty questions on two engines is enough. For a share you want to quote, run forty to a hundred questions and repeat them, which is the level at which the figure holds still between runs.

Why a company that sells the alternative published this number

AI Knows Us sells exactly what this measurement points at: running buyer questions across ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews, finding which sources decided the answer, and writing the pages that get read. A finding that says review directories are 0.8 per cent is commercially convenient for us, which is a good reason to publish the method, the dates and the counts alongside it so you can take it apart. The same run gave our own domain 3 citations in 597 readings, which is the worst number in the dataset and belongs to the company that collected it. Count your own twenty questions and you will have your own version of both figures by this evening.