Review sites reach ChatGPT's answers through retrieval, and what retrieval can read is page text, not a star average. When ChatGPT answers a shortlist question with search on, it runs one or more targeted queries and builds the answer from the pages that come back, and in most categories a G2, Capterra, or Trustpilot category page is among them. More than 4 in 10 AI citations point to sources brands do not control, mostly editorial coverage and platforms like Reddit and YouTube, and review platforms add another third-party record that buyers and engines read. A profile the engine can retrieve and quote for the question asked is evidence; a profile it never retrieves is a rating nobody reads.
So the answer is yes, with a caveat that changes the budget: the public evidence is correlational. The vendors with the most reviews are usually also the most established and best ranked, so the model would name them with or without the profile, and the controlled experiments that exist test on-page changes, where C-SEO Bench found most rewrite methods largely ineffective and often harmful to citation rank.
The stakes justify the care. G2's research found AI chatbots and review sites are now the two biggest influences on B2B shortlists, 51% of software buyers now start research with an AI chatbot more often than with Google, and a third of buyers said an AI chatbot led them to choose a vendor they hadn't considered.
Does ChatGPT actually read G2, Capterra, and Trustpilot when it recommends software?
Yes, through two paths. The first is live retrieval: for current commercial questions ChatGPT searches and builds the answer from the pages it reads. Seer found 87% of SearchGPT citations matched Bing's top results in a 500-citation sample, and in most categories a review-platform page holds one of those positions for shortlist queries. That correspondence is not universal: Ahrefs found 38% of AI Overview citations rank in the organic top 10 across 863,000 SERPs in March 2026, so ranking is a route into the answer and not a reservation.
The second path is memory. OpenAI describes its training data as drawn from three primary source classes, the first being publicly available internet content, which includes years of indexed review pages. A 2026 analysis of brand dynamics in LLM recommendation systems, which gave the models product details directly, found well-known brands were recommended every time only when the specifications were identical; a competitor's rating edge of under 0.1 stars erased that lead. Whether those pages build a vendor's prior is untested, which is the gap this article keeps running into.
No public evidence shows the numeric rating itself is an input to either path. What is demonstrably retrievable is the page text: the category, the description, the reviewers' use cases, the integrations, and the pricing fields. AirOps' March 2026 analysis of 548,534 pages ChatGPT retrieved found only 15% were cited in the final answer, and a review profile faces the same cut as any other page.
Is the correlation between review profiles and AI mentions proof that reviews cause recommendations?
No. The argument usually rests on a count: the tools ChatGPT names in a category nearly all have G2 or Capterra profiles, therefore the profile caused the naming. Whatever that count shows in your category, the inference fails on three grounds.
First, confounding. Vendors with deep review histories usually also have the most third-party coverage and the strongest rankings, and buyers already know them: 6sense found nearly four out of five vendors are named to the shortlist on day one. An engine that names those vendors is reporting the same maturity the reviews report, and no count of profiles can tell the two apart.
Second, variance. SparkToro's research found AI engines are highly inconsistent when recommending brands, and Search Engine Land's coverage of the same study notes the recommendation lists rarely repeat exactly. A 2026 variance-components study found run-to-run noise large enough to swamp real differences in small samples. A before-and-after on one prompt, taken the week your G2 profile went live, is a coin flip dressed as a result, which is why Ronald Sielinski's "Quantifying Uncertainty in AI Visibility" insists on confidence intervals.
Third, the controlled evidence tests something else. The GEO study (Aggarwal et al., KDD 2024) lifted citation visibility by roughly 30 to 40% in its benchmark by adding statistics, quotations, and citations to pages, and C-SEO Bench tested rewrite methods across two tasks and six domains and found most of them largely ineffective and often harmful to citation rank, with traditional retrieval-side improvements doing more. No public experiment adds a review profile to an otherwise identical vendor and measures the recommendation rate, so review sites remain an evidence surface with a plausible mechanism and no isolated effect size.
What on a review profile actually becomes evidence in an answer?
The parts an engine can quote for the question the buyer asked. A shortlist question carries constraints: team size, use case, required integration, budget, region. The profile that gets used is the one whose description, category tags, integration list, and pricing fields state those constraints as plain text, so "pipeline and forecasting for sales teams of 5 to 25 people" is retrievable for that question and "next-generation CRM for modern teams" is not.
Reviewer text matters more than the review count. Reviews that name the job to be done and the outcome give the model something to extract. Google's ecommerce guidance for writing product reviews lists 14 recommendations, and almost all of them ask for specific, evidence-bearing detail, the same property that makes a G2 review quotable in an answer.
Persuasive framing on the profile runs the wrong way. Scarcity and exclusivity framing measurably reduces how often an LLM recommends a product, in an experiment run on three sets of 10 fictitious products, which removes brand familiarity. Consistency is an entity problem: a vendor listed under one name on G2 and a legacy name on Capterra splits into two weak strings, which makes it harder for readers and engines to attribute reviews to the right company.
Concentrate. Reviews spread thin across six platforms are thinner evidence than the same reviews concentrated on one, because each profile is retrieved on its own merits. A review-site page is also a comparative surface, so when the engine cites the G2 category page it reads your row and your competitors' rows in the same pass. The profile buys you presence in an answer whose shape you do not control, which is the argument for also spending on sources where you are the only vendor named.
How do you get reviews without breaking platform rules or FTC guidance?
Ask real customers at real moments, and never pay for the rating. Requests convert around documented success: onboarding completion, a first measurable outcome, a renewal, an expansion. Send a direct link, follow up once a week later, then stop, because sequences that keep pushing turn satisfied customers into two-star reviews.
Do not buy reviews, reward specific ratings, or fabricate reviewer identities. Each violates platform terms, and fake reviews and undisclosed incentives sit inside the FTC's advertising guidance. Forty reviews posted in one week by accounts with no history reads as manipulation, and a flagged profile is worth less than an empty one.
Forecast from your own funnel. A company with 100 active customers can estimate a quarter's review volume from eligible trigger events, request delivery, response rate, and platform acceptance. Any "50 reviews in 60 days" target not derived from your own event volume is a wish, and no public study establishes 50 as the number an engine wants.
How do you measure whether review-site presence changed what ChatGPT says?
Establish a base rate before you touch a profile, because one answer proves nothing. Write the five to ten questions a buyer asks before shortlisting, ask each in two or three phrasings on ChatGPT and at least one other engine, once a week for a few weeks, and record who was mentioned, cited, and recommended, plus which sources carried the answer.
Classify what you find by rung. The IAB's 2026 guidelines on measuring visibility in the AI era organize the metrics into four groups for brands, Presence, Prominence, Portrayal, and Persuasion, and a review profile can move one without moving the next. Being listed on a category page the engine cites is presence. Whether the answer then describes you accurately is portrayal.
Re-run after the change and compare distributions. Moving from 1 of 10 runs to 4 of 10, across two engines, on questions where a review page sat among the citations, is evidence; a screenshot is an anecdote. Keep the competitor columns too, since a rival's new profile can move your numbers as much as your own.
How does Trovance show whether review sites are carrying the answers in your category?
Trovance runs the measurement above as a standing system instead of an afternoon exercise. You define the buyer questions that matter, Trovance sweeps them across engines every week as tracked questions, and every answer is preserved as a snapshot with its full context: mentions, citations, recommendations, and the sources that carried the answer. That record lets you observe, for your category rather than an industry average, whether review pages are among the sources at all.
Share of answer turns the record into a diagnosis. When a review-site page carries an answer that omits you, the recommended action names the source that carried the answer without you, so the profile work has a target. When your profile is cited and the recommendation still goes to a competitor, the snapshot shows what was quoted from their page and not from yours, which is a claims-and-proof problem. The Brain keeps the claims you are entitled to make with the proof behind each, so a review-derived claim can carry its source and date.
When the missing asset is on your own site, drafts are written in your voice from your voice guidance and research sources, a person edits and approves them before anything publishes, and the next analysis cycle verifies whether the answer moved.
What Trovance will not do is collect reviews for you, post to review platforms, or promise that a completed profile changes the recommendation. No honest system can promise that, because the engines are probabilistic and your competitors are filling in their profiles too. What it does instead is close the loop: after the profile work goes live, the next cycle reruns the same questions and compares distributions, so you learn whether review-site evidence moved the answer.
What should you do this week?
Measure first. Run your five to ten buyer questions in two or three phrasings on two engines for a few weeks and record who was cited, so you know whether review sites carry answers in your category. Then fix the profile as a page: one name everywhere, a description in buyer phrasing, every structured field filled with something you can defend. Start the request flow from your real trigger events with no incentive attached.
Be honest about the timeline and the ceiling. Reviews accrue at the pace of your customer success, retrieval catches up on its own schedule, and every measurement carries variance, so anyone promising a recommendation in exchange for a review count is selling against the evidence. If review sites were never among the citations for your questions, put the budget into the sources that were. If you want the base rate without the manual runs, start a free Trovance trial and see which sources are carrying the answers in your category.
Earn the off-site half of the evidence
Where AI citations come from across industries - how much of the answer sits off your domain.
How to earn third-party AI citations - the sources beyond review sites that carry answers.
How Reddit shapes AI recommendations - the community half of off-site evidence.
Social proof is an attack surface, not a GEO tactic - why manufactured proof backfires.
Measure before you spend
Why ChatGPT recommends your competitor - the five failure modes.
The 10-minute audit of what ChatGPT tells buyers about you - the fastest first reading.
There is no such thing as an AI visibility score - why one number hides the rungs.
The most-quoted study in AI search, quoted correctly - what the GEO benchmark actually tested.
