Blog · Third-party sources

Review sites, G2 and what AI repeats about your product

Published 8 October 2026 · Updated 12 October 2026 · 8 min read

Review sites, G2 and what AI repeats about your product
Photograph: Cheung Yin / unsplash

The short answer: In the audits we have run, review directories are a frequent source when an engine describes a product, and the pages we see lifted are usually category and comparison pages, rarely a single review. We read citations in answers, not engine internals, so take that as a sample. You cannot edit the sentence an engine writes. You can change the material it is built from: the profile, the recency of reviews, and whether anything else repeats the claim.

That is the whole mechanism, and almost every mistake in this area comes from skipping the middle step: assuming the engine read your reviews, when it mostly read a roll-up of them.

The directory is quoted because it is not you

An assistant writing a shortlist has an awkward job. Every vendor site in its candidate set says that vendor is the best choice. A directory page says something different in kind: it carries an aggregate, a rank within a category, and language that was not written by the company being described. For a model optimising for a defensible answer, that is cheap corroboration.

The effect compounds on comparison pages. A page titled around a category, listing eight products with ratings beside each, answers a shortlist prompt almost word for word. It needs no synthesis. Work on this blog has covered the pattern from the other side in The listicles you are missing from, and the underlying preference is the same one described in The third-party sources AI engines keep citing.

What this means in practice is unflattering. Your category page on a review directory may carry more weight in an answer than your own product page, and you do not control a single sentence on it.

Run the robots check before you believe anything

Here is a check you can run this afternoon. What follows is one fetch of each file, made on 12 October 2026 from a single Indian network, and a robots.txt can change the next day, so read it as a worked example rather than a standing description of these sites.

Fetch the robots.txt of every directory your category lives on and look for the AI user agents by name. G2's file groups GPTBot, ClaudeBot, Applebot-Extended, Meta-ExternalAgent, Amazonbot and CCBot into one block. That block repeats every path rule from the wildcard group and then adds thirteen more, among them Disallow: /products/*/reviews/*, the three-way comparison paths /compare/*-vs-*-vs-*, and the /de/, /fr/, /es/, /pt/ and /it/ locale trees. So the named AI agents are told to stay off the product review pages that the wildcard group may read. Notably absent from the named group are OAI-SearchBot and PerplexityBot, which appear nowhere in the file and therefore fall through to the wildcard rules.

That distinction matters more than it looks, because OpenAI's own crawler documentation states that each of its agents is configured independently: OAI-SearchBot surfaces pages in ChatGPT's search features, while GPTBot collects training data. A site can be absent from one and present in the other.

Trustpilot takes a different route. Its file names OAI-SearchBot, PerplexityBot, GPTBot, Google-Extended, ClaudeBot and CCBot, gives each of them the rule set it gives Googlebot, and that rule set disallows /reviews/ and /product-reviews/ along with every filter parameter. Bytespider gets a flat disallow. Capterra and Software Advice both list Bytespider and DeepSeekBot in a long group that is disallowed outright, and both declare a non-standard LLMS TXT: line inside robots.txt. The robots exclusion standard, RFC 9309, defines no such field, so a compliant parser is not required to read it. We have not seen evidence either way that engines read it. PeerSpot goes the other way entirely, with a commented allow list for six AI agents at the top of the file.

Then there is the layer underneath robots. On 12 October 2026 we made one request each to two of these domains with an ordinary desktop Chrome user-agent string, no JavaScript and no logged-in session: a G2 product review URL returned HTTP 403 with a 1,704 byte body, and a Capterra category page returned HTTP 403 with a 5,485 byte Cloudflare body. That is two requests from one network on one day, not a census, and an engine's crawler comes from different addresses and may be treated differently. A path can be permitted in robots.txt and still refused at the edge. Allowed and fetchable are two separate questions, and only the second one decides whether a passage reaches an answer.

Three things an engine actually lifts from a profile

Watch which fragments survive into answers and a pattern shows up.

The aggregate number. A rating and a count, usually together. Schema.org defines ratingValue alongside ratingCount and reviewCount as separate properties, and engines quote them as a pair because one without the other is meaningless. A 4.8 from nine reviews and a 4.3 from four thousand are different facts.

The one-line category placement. Phrases of the "leader in" or "top rated for" shape, lifted from a badge or a grid position. These are the fragments that recur most often in the answers we read back during audits, and the most dangerous, because a placement from two quarters ago reads as present tense in an answer with no date attached.

A complaint cluster. Not one review. A theme repeated across many, which the directory itself often summarises in a pros-and-cons block. If eleven people wrote that onboarding took six weeks, that sentence is now part of your product description in a dozen assistants.

Pricing is the special case, because directories carry a pricing field that goes stale quietly and then gets quoted with confidence. We have written separately about Pricing pages and what AI says your product costs; the directory copy of your price list is the second place to check, and in the audits we have run it has been out of date more often than the vendor's own page. We have not counted that across a fixed sample, so take it as a habit worth checking rather than a rate.

How to change the sentence without being able to edit it

You cannot file a correction with a model. The practical route is slower and it works.

Start with the fields you do control. Most directories let a claimed vendor edit the description, the category tags, the feature list, the integrations and the pricing tiers. Those fields are what feed the roll-up text engines quote. A claimed profile with a current description beats an unclaimed one with four-year-old marketing copy, and the edit takes an afternoon.

Next, attack recency rather than volume. The directories we work in sort and surface recent reviews prominently, and the roll-up text an engine quotes tends to follow what is on the page now, so a fresh run of reviews this quarter tends to move the quoted sentence further than a much larger pile from three years ago. We cannot see how far back any directory's own summary looks, because none of them publish that, so this is a working rule from audits rather than a documented behaviour. Ask customers who just renewed, not customers who just signed.

Third, make the claim resolvable on your own property too. When your site says the same thing as the directory, in markup an engine can parse, the engine has two agreeing sources instead of one. That is the argument for Product schema for AI shopping answers. Be careful with one detail here: Google's review snippet policy states that when the entity being reviewed controls the reviews about itself, pages using Organization or LocalBusiness markup are ineligible for the star feature. Self-hosted testimonials marked up as AggregateRating are the classic own goal.

Then measure it the only way that holds up. Ask the same ten buying questions across several engines, record which sources were named, and re-run after the profile edit lands. The self-check walks through the mechanics if you have not run a prompt set before.

What we are still measuring

This post establishes less than it may read like. The robots and HTTP findings are single fetches of five files and two pages on 12 October 2026 from one network, so they describe those files on that day and nothing more durable. We do not know how long a directory edit takes to reach an answer, because the lag combines the directory's own re-render, the crawler's return visit and the engine's index refresh, and we have seen it range from under a fortnight to more than a quarter on the same site. We have not counted how often a directory page is cited versus a vendor page, so the claim that directories are quoted frequently is a pattern across our audits, not a measured share. And we cannot tell, from the outside, whether an engine reached a page through a live fetch or through a search index that cached it months ago: a disallow in robots.txt does not prove nothing was read, it only tells you what was asked for. The robots check tells you what is permitted, not what reached the answer.

FAQ

Does paying a review site improve what AI says about my product? Paid placement buys badges and category positioning, which can end up in quoted text. It does not change the review content itself, and the aggregate and the complaint clusters are doing most of the work in the answers we read.

Should I ask a model to correct a wrong claim about my product? That correction lives in one conversation and does not persist. Change the sources instead, starting with the directory profile fields and your own markup, then re-run the prompts in a fresh session to check.

Which review site matters most for my category? Run the prompts and read the citations rather than guessing. Category concentration is real, and in our audits the dominant directory differs sharply between developer tooling, HR software and anything bought by a procurement committee.

My review profile is thin. Is that worse than a bad rating? Usually, yes. A low count gives an engine nothing to aggregate and it tends to reach for a competitor profile that has more signal, which removes you from the comparison entirely.

Can I see which third-party pages an audit checks? The sample audit report shows the shape of the offsite section, including how a cited directory page is traced back to the prompt that surfaced it.

Sources

All URLs below were opened on 12 October 2026, and the robots.txt files were read on that date.

Related reading

The third-party sources AI engines keep citingThe listicles you are missing fromProduct schema for AI shopping answersPricing pages and what AI says your product costsWikipedia, and the brand facts engines inheritYouTube, transcripts and video citations