Why AI Answers Cite G2 and Trustpilot, Not Your Homepage

Ask an assistant to recommend a supplier and watch what it reaches for. Not your homepage. It pulls from listings, comparison pages, forum threads and review platforms, and it does this even when your own site is crawlable and says all the right things about you.

This is not a slight. Your landing page is the one document on the web guaranteed to be biased in your favour, and a system trying to answer "which vendor is best for this" has to weigh it accordingly. When a person asks a friend for a recommendation, the friend does not read out the brochure.

Your own site is a weak source about you

Your site still matters. It is where a model learns what your product is: the category, the pricing model, the integrations, the use cases. Bury that under positioning language and a model has nothing concrete to reuse.

But there is a class of question your site cannot answer, and it is the class that decides a purchase. Is it any good. Who is it right for. Where does it fall down. Every vendor claims to be excellent, so the claim carries no information. A judgement is only worth quoting if the party making it had the option of saying something unflattering.

Review platforms are that party. G2, Trustpilot, Capterra and their peers exist to host verdicts the vendor did not write, and that is what makes their pages evidence rather than marketing.

What a review listing looks like to a machine

Set the trust question aside, because there is a more mechanical reason these pages get used: they are shaped like an answer.

A review listing is a page about exactly one entity. The product has a name, a vendor, a category, a description, stated features, the things it integrates with, and a body of opinions attached to it. There is no ambiguity about what the page is about, which is rarely true of a company website, where one URL is trying to be a pitch, a hub and a navigation surface at once.

The listing is also organised by category and by use case, and that is the crucial detail, because it mirrors the way the question arrives. Buyers do not ask "tell me about brands". They ask for the best tool in a category, for a particular kind of company, for a particular job. A page that already sorts products by category and who it suits has done, in advance, the work the assistant would otherwise improvise.

And it is comparative: your product sits next to the alternatives, with the differences described by people who chose between them. A model summarising "the leading options for X" has, on one page, most of the raw material for that sentence. The pages are updated as new reviews arrive, too. Current, structured, unambiguous about its subject, not written by the party being judged: it would be strange if assistants ignored a source like that.

The checks you can run today

Most failure modes here are boring and fixable, and none of what follows needs a budget.

Does your listing exist at all

Search for your product on the major platforms in your category. Plenty of companies discover they are not there, or that an abandoned entry exists, created by someone who has since left. If there is no page, an assistant looking for third party evidence about you finds nothing, and quotes whatever it does find instead.

Is the listing claimed

An unclaimed listing is usually a stub: a name, a scraped fragment of a description, no screenshots, no feature list, no reply to any review. Claiming it is free on most platforms and takes an afternoon. It hands you control over the parts of the page that describe the product, which are exactly the parts a model will lift.

Are you in the right category

This is the most damaging thing to get wrong, and almost nobody audits it. Categories are the taxonomy through which the recommendation question gets answered. Filed under the wrong one, you get two bad outcomes: you surface where you will be dismissed as a poor fit, and you are absent from the recommendations you would have won.

Go and look at which category you are in, and at which products sit in it with you. If the neighbours are not the ones your buyers weigh you against, the filing is wrong, and most platforms let you request a change.

Does the description say what you do, plainly

Read your listing description as if you had never heard of the company. Does it state, in ordinary words, what the product does, who it is for and what problem it solves? Or does it announce that you are a leading platform empowering teams to unlock their potential? The second version gives a model nothing it can put in a sentence, and nothing gets recommended that cannot be described.

The same goes for the structured fields nobody fills in: features, integrations, pricing model, the segments you serve. Those are machine readable, and they decide whether you are retrievable for "tool for small teams that integrates with X".

Where the honest line runs

Asking your customers for reviews is legitimate, and if you have never asked, that is why you have so few. Ask the ones who already told you they were happy, while the value is fresh, and do not tell them what to write.

Buying reviews is not legitimate, and neither is writing them yourself, having your staff write them, or making a discount conditional on a positive one. Platforms invest heavily in detecting exactly this, because their whole business rests on being the source that cannot be bought. A listing carrying a fraud notice is worse than no listing: you have handed every retrieval system a documented reason to distrust you.

And nobody, us included, can promise that a well maintained listing will get you cited. The mechanism is plausible and these pages are the kind of source these systems reach for, but there is no submission form and no lever. What you can do is remove the reasons you are being skipped. That is a smaller claim, and a true one.

How you would actually know

All of this stays a hypothesis about your brand until you look at what assistants actually cite when they talk about your category. That is a measurement problem, not a content problem. By hand: ask the questions your buyers ask, in each language you sell in, and read which sources come back. When the matrix of prompts, platforms and languages grows past what a person can run by hand, that is where PSentry comes in: it runs your prompt set across ChatGPT, Claude, Gemini and Perplexity, reports where you are named and where you are cited, and shows the sources behind it. Think of it the same way as the listings themselves: it reads what is already there and reports it back, on a fixed cadence twice a month rather than by the hour. It has no pen to add a line to a review, and no promise attached to what it finds.

Do not hide the bad reviews. Read them.

Here is the unwelcome part. Once a listing is feeding recommendations, an assistant will sometimes summarise your weaknesses, because the weaknesses are on the page. Onboarding is confusing. Support is slow. The reporting is thin.

The instinct is to treat this as a visibility problem and go looking for a way to suppress the source. It is not one. If several unrelated customers wrote the same complaint and a machine noticed the pattern, the pattern is real. The assistant did not invent it and did not editorialise. It read what your customers said and compressed it.

So the fix is the product, or the onboarding, or the support queue, and the review page is only the instrument that made it audible. Reply in public, without defensiveness. Fix the recurring thing. Treating the mirror as the problem is how a company spends a year optimising its way around a truth its customers had already written down for it.

Frequently Asked Questions

Which review platforms matter for my category?

The ones your buyers actually consult, which vary by industry. Business software clusters on a handful of well known platforms, while consumer services, hospitality and local trades live somewhere else. A proxy: ask the comparison question your buyer would ask, and see which listing sites come back.

Do I need a lot of reviews before this does anything?

Volume helps, but presence and correct classification come first. A listing that does not exist, or that files you under the wrong category, cannot be rescued by more reviews.

Can I get an unfair review removed?

Platforms have dispute processes, and reviews that break their policies (fabricated, defamatory, written by someone who was never a customer) can be taken down. A review that is merely negative but honest will stay, and should. Replying in public is usually the better move, because the reply becomes part of the page a model reads.

Are review platforms the only third party source that counts?

No. Documentation, forum threads, comparison articles written by other people, marketplace listings and press coverage all serve the same function: a description of you that you did not write. Review platforms are just the most structured and the most explicitly comparative of them.

My listing is fine and I am still not cited. What now?

Then the listing was not the constraint. Look at which sources are actually cited for your category, and in which language, because visibility does not carry across markets. Absence in one language and presence in another is common, and it points at the web in that language rather than at your listing.