Do Review Sites Affect What AI Recommends?

Do assistants actually cite review sites?

Yes, reliably, and less than the pitch decks imply. No assistant operator publishes citation shares by domain, so this rests on vendor observation. The largest published dataset is Profound's, released 5 June 2025 and updated August 2025, covering 680 million citations across ChatGPT, Google AI Overviews and Perplexity from August 2024 to June 2025. Profound sells answer-engine optimisation software and has an obvious interest in the finding that AI citations concentrate on a short list of influenceable domains; the numbers are reported here with that stake named.

What they show is a category that is always present and never large. G2 is ChatGPT's fourth most-cited domain at 1.1% of all citations, behind Wikipedia at 7.8% and Reddit at 1.8% and level with Forbes. Gartner appears fifth in Google AI Overviews at 0.7%. Perplexity's top ten contains four rating destinations at once — Gartner at 1.0%, Yelp at 0.8%, G2 at 0.6% and TripAdvisor at 0.6%.

Two cautions before anyone builds a budget on this. Neither Capterra nor TrustRadius nor Trustpilot appears in any published top ten, which means their share is below the cut-off and unmeasured — not that it is nil. And a share of all citations across a mixed prompt set says nothing about the share on the prompts that matter to a B2B software buyer, which is the population any vendor actually cares about. The same denominator problem runs through which sources AI assistants actually cite.

Who owns the review corpus now?

This is the part most commentary on the subject has wrong, including our own earlier framing of these platforms as four independent competitors.

They are not four. G2 states in its own newsroom that its marketplace network “now includes Capterra, Software Advice, and GetApp” — the three properties that previously traded as Gartner Digital Markets. On G2's own figures the combined network “reaches more than 200 million buyers globally each year and features 6 million verified reviews across 200,000+ software products in 2,000+ categories”. TrustRadius, the fourth name usually listed, was reported acquired by the market-intelligence firm HG Insights in 2025.

Take those together and the practical position is that three of the four destinations a B2B buyer is told to consult sit under one owner, and the fourth sits under a data company that sells intelligence to the same vendors it rates. Trustpilot remains independent and separately listed, but it is a consumer reviews platform: it rates the company, not the software category, and it does not produce “best CRM for mid-market” style rankings at all.

Why this matters for an AI answer is mechanical rather than conspiratorial. When an assistant assembles a “best X for Y” response from several review sources, the apparent corroboration — four sites agreeing — may be one editorial methodology, one moderation team and one incentive policy appearing four times. Diversity of citation is not diversity of evidence. The same trap applies to syndicated coverage, which we covered in how third-party authority actually works for SaaS.

How the reviews get there: gating, verification and incentives

The platforms publish their rules, and reading them side by side is the most useful thing a buyer can do, because the corpora are built to different specifications and an assistant reading both is averaging incompatible things.

G2: identity-gated, incentive-permitting, aggressively scoped

G2's Community Guidelines require a reviewer to validate identity through a LinkedIn account, a verified business email, or a seller-authenticated route. Employees of the vendor and of direct competitors may not review. Reviews based on an experience more than two years old are not allowed, “as B2B technology can change drastically in a relatively short amount of time”. Every review passes automatic filtering and then a manual check before publication.

Two exclusions matter more than they look. Business-partner reviews — resellers and similar — are permitted but “do not count toward the G2 Score” and carry a label. Guest-user reviews, since June 2024, likewise do not count toward the score. So the visible review wall and the scored corpus are different populations, and a reader who counts stars on the page is not seeing what the ranking is computed from.

On incentives, G2 is permissive and explicit: it “will occasionally offer incentives (such as gift cards or a donation) for honest reviews”, eligibility “is never based on the opinions, positive or negative, expressed in the review”, confirmed incentivized reviews are tagged as such, and there is a hard ceiling — “G2 limits the value of any incentive offered in exchange for a review to $100 USD”, covering cash, gift cards, swag, tokens, credits and subscriptions.

G2 also has the field's most developed rule on machine-written reviews. AI drafting assistance is permitted only if the reviewer is LinkedIn Verified, is on a seller-provided list, or is submitting in-app; unverified reviewers may use Grammarly or Google Translate but may not have an AI generate the review. Allowing AI to “fabricate experiences, results, or product claims” is prohibited outright. Whether that rule is enforceable at scale is a separate question, and G2 publishes no detection rate.

Trustpilot: open, uninvited, no incentives at all

Trustpilot's Guidelines for Reviewers, version 2 dated May 2025, describe a deliberately opposite design. Anyone over 18 with a user account and a genuine recent experience may review, without an invitation; “recent” typically means the last twelve months. An experience need not be a purchase — “a phone call, email, online chat, or visit to an online store, even if you decided not to buy anything” qualifies.

And the incentive rule is a flat prohibition: “We don't allow incentivized reviews on Trustpilot. This means you're not eligible to write a review if you've received or been offered an incentive in connection with writing or editing a review”, including discounts, promo codes, prize-draw entries, refunds and freebies. People with a special relationship to the business — owners, family, staff, shareholders, competitors — are excluded, to prevent bias “both positive and negative”.

The consequence is worth stating plainly: a campaign that is fully compliant on G2, with a $50 gift card and an “Incentivized” tag, is a guidelines violation if the same reviewer is pointed at Trustpilot. Any agency selling “review generation” as one service across both platforms is selling two different products, one of which breaks the rules. The disclosure obligations that sit underneath all of this are covered in what the FTC requires for paid placement, and G2's business guidelines cite the FTC's review rules directly.

Are the assistants reading these sites, or licensing them?

Both, and the licensing part is newer and less discussed.

G2's Guidelines for Businesses prohibit anyone from using the site or its contents “for any commercial purpose without our express written permission”, and from accessing, monitoring or copying its content “using any robot, spider, scraper, or other automated means or any manual process for any purpose without our express written permission”. Bypassing robot exclusion headers is separately prohibited. On the face of the terms, an AI company that crawls G2 to build a recommendation product is outside them.

What exists instead is a commercial channel. G2 has announced that it is part of OpenAI's launch of the Data agent in ChatGPT Work, bringing “trusted B2B software market data into ChatGPT Work”, with OpenAI's general manager for technology, Arpan Shah, quoted saying that the integration “lets people explore those signals and build their own dashboards using natural language”. G2 also operates a G2 MCP connector that puts its buyer-intent and verified-review data into AI workflows, and says “more than 1,200 users across 550+ organizations” are already querying it.

Note what G2 is selling there, because it is a tell. Among the things a vendor can ask its data about is “how competitive your traffic and answer engine optimization (AEO) visibility stands in comparison to your peers”. The review platform is now also selling AI-visibility analytics to the companies it rates. That is not a scandal, but it is a conflict worth knowing about before treating a G2 report as a neutral reading of the market.

The structural point is the same one that governs Reddit's role in AI answers, which we set out in do Reddit threads show up in AI answers: the corpora that assistants lean on hardest are increasingly ones that are contractually closed to open crawling and commercially open to a small number of buyers. A citation from such a source reflects a supply agreement as much as an editorial judgement.

Does any of this move a “best X for Y” answer?

Here the evidence runs out, and the honest thing is to say so rather than to bridge the gap with an inference.

What is established: assistants cite review domains, at roughly 1% of citations each; those domains hold millions of structured, verified, category-tagged opinions in exactly the shape a comparison question needs; and at least one of them is being licensed directly into an assistant.

What is not established: that adding reviews to a product's profile changes what an assistant says about it. We could find no controlled study — vendor, academic or platform — that varies a product's review count, rating or recency and measures the change in a “best X for Y” response. Every argument in the market for review-generation as an AI-visibility tactic is an extrapolation from a citation share to a causal effect, and that step has not been tested by anyone who published the result.

Three things follow for a buyer.

Treat the category page, not the review, as the asset. What an assistant is most likely to retrieve is a ranked category list, not an individual testimonial. Being absent from the category is a visible, checkable problem; being the 41st review on a profile is not obviously anything. The mechanics of how those lists get assembled are the subject of how AI assistants recommend companies.

Read the gating rules before you buy the campaign. Business-partner and guest-user reviews do not count toward the G2 Score. Reviews of experiences older than two years are not accepted. An incentive over $100 is a violation. A campaign that produces volume outside those boundaries produces a wall of text that the ranking ignores.

Price the badge honestly. G2's rules require a paid Content Subscription before a vendor may republish numeric rankings from a G2 report, and require attribution, dating and a backlink for live Grid placements. The award is earned from reviews; the right to advertise it is a licence you rent. That distinction is the same one that governs paid media evidence generally, discussed in does media coverage affect B2B buying.

Media Morpheus sells media and podcast placement, not review generation, and nothing here would support adding it. The finding that should reach a founder is narrower and more useful than a tactic: the review corpus matters to AI answers, it is owned by fewer companies than it appears to be, its rules differ enough that one campaign cannot be compliant across it, and the causal claim attached to it has never been measured.

Methodology and limitations

Citation shares are from “AI Platform Citation Patterns: How ChatGPT, Google AI Overviews, and Perplexity Source Information” by Nick Lafferty, Profound, published 5 June 2025 and updated August 2025, covering 680 million citations observed August 2024 to June 2025. Profound sells answer-engine optimisation software; the prompt set is not published, so the query mix is unknown and cannot be matched to a B2B software buyer's queries; the study reports only each platform's top ten domains, so Capterra, TrustRadius and Trustpilot are below the reporting threshold and their shares are unknown rather than zero; and the products change continuously, so the figures are a period observation. Platform rules are quoted from the publishers' own current policy pages, read on 24 September 2026: G2's Community Guidelines, covering guidelines for reviewers and for businesses, and Trustpilot's Guidelines for Reviewers, version 2, dated May 2025. Neither platform publishes an enforcement or detection rate, so the effectiveness of these rules is a matter of the publishers' assertion and is not verifiable from outside; both have a commercial interest in their corpora being regarded as trustworthy. Ownership: that G2's network now includes Capterra, Software Advice and GetApp is taken from G2's own newsroom statement and is therefore primary; the completion date of that transaction and the reported acquisition of TrustRadius by HG Insights in 2025 are reported rather than verified here, because we could not open a primary filing or announcement for either, and they should be treated as such. The G2 statements about the OpenAI Data agent in ChatGPT Work, the G2 MCP connector, the 6 million verified reviews and the 1,200 users across 550+ organizations are G2's own marketing claims about its own products, unaudited and uncorroborated. We deliberately did not use the Trustpilot commissioned study produced with the agency SEER Interactive on review platforms and AI visibility: it is a vendor study about its own publisher's platform, and we did not open it, so none of its figures appear above. Finally, and most importantly, we found no study of any kind that manipulates a product's reviews and measures the resulting change in an AI recommendation; the central causal question in this guide is unanswered in the public record, and we have left it unanswered rather than inferred.

Conclusion

They are in the mix, and their influence is smaller and stranger than it is sold as. Review destinations sit inside every major assistant's top ten cited domains at roughly 1% of citations each — always present, never dominant. The category is also far more concentrated than it looks: G2 now carries Capterra, Software Advice and GetApp, leaving Trustpilot as the significant independent and a consumer platform at that. The corpus is not an open public record either; G2's terms forbid scraping it, while G2 sells the same data into ChatGPT Work. And the question a founder actually wants answered — whether more reviews produce a better AI recommendation — has not been tested by anyone who published the answer. On present evidence the defensible move is to be correctly listed and correctly categorised on the platforms, to follow their gating rules rather than their sales teams, and to disbelieve anyone who quotes a citation share as though it were a conversion rate.

Frequently Asked Questions

Sources

1. vendor Lafferty, N., 2025. AI Platform Citation Patterns: How ChatGPT, Google AI Overviews, and Perplexity Source Information. Profound, 5 June 2025, updated August 2025. 680 million citations; Profound sells answer-engine optimisation software.
2. platform policy G2. Community Guidelines. Guidelines for reviewers and for businesses, review moderation labels, the $100 incentive cap, the generative-AI rule and the scraping prohibition. Read 24 September 2026.
3. platform policy Trustpilot. Guidelines for Reviewers. Version 2, May 2025. Prohibition on incentivized reviews; twelve-month recency; open, uninvited reviewing.
4. operator Abel, G. G2 Brings Verified Software Market Data to OpenAI's ChatGPT Work. G2 newsroom. Source for the marketplace network including Capterra, Software Advice and GetApp; the 6 million verified reviews figure; the G2 MCP adoption figures; and the quote from OpenAI's Arpan Shah. Marketing claims by G2 about G2.
5. research Chapekis, A. and Lieb, A., 2025. Google users are less likely to click on links when an AI summary appears in the results. Pew Research Center, 22 July 2025. The only non-commercial measurement of AI summary sourcing, used here for context on citation mix.