Do AI assistants read your reviews? G2, Trustpilot and Reddit in AI answers

Retrieval mostly follows conventional search. When an assistant with live web search handles a question like 'best CRM for a small accountancy firm', it fetches whatever ranks for that query, and for software that usually means G2 and Capterra category or comparison pages. Those pages were built to rank for buying questions, which is exactly why they keep appearing in AI citations.

For consumer products and services, Trustpilot carries similar weight, particularly for UK and European brands. Reddit cuts across every category: threads read like candid buyer conversations, and Reddit's licensing deals with Google and OpenAI mean its content flows into training data as well as live retrieval. A complaint thread from eighteen months ago can reach an answer through either route.

Both routes are error-prone. In our July 2026 test of 30 brands, assistants misdescribed 27 of the 30 (90 percent) when answering from memory, and 13 of the 30 (43 percent) even with live search switched on. Review content is one of the main inputs behind those descriptions, so what your review pages say matters whether or not they are fetched fresh.

A language model builds answers out of text, and an aggregate rating gives it almost nothing to work with. A 4.5 average says nothing about who the product suits or what it replaced, and mature products tend to bunch together in the mid fours anyway, so the number barely separates you from the competitors listed beside you.

Written comparative reviews are different in kind. 'We moved from X to Y because the reporting worked for a regulated industry' is a miniature answer to a buying question, and when an assistant decides which brands to recommend, sentences like that are what it synthesises from. The review has done the assistant's work in advance, so it gets used.

Review platforms show recent reviews first, and a crawler sees what the page shows. If your latest G2 review is two years old, an assistant reading that page describes the product as it stood two years ago, bugs included, and presents that stale description as if it were current.

Specificity governs whether a review is quotable at all. Buying questions arrive narrow ('for a UK charity', 'for a warehouse team of 40'), and the review most likely to be quoted is the one whose details match the question. A review that names an industry and a task can be lifted almost word for word into an answer. 'Great product, five stars' cannot.

Ask happy customers for written reviews and make the ask concrete. Not 'please leave us a review' but 'would you write a couple of sentences on what you use us for and what you switched from'. You are not scripting the review, which every platform prohibits; you are prompting for the detail that makes a review quotable.

Respond to negative reviews with facts rather than apology boilerplate, because responses sit on the same page and get read too. A dated, factual reply ('the export bug described here was fixed in the March 2026 release') gives an assistant a correction to weigh against the complaint. Silence leaves the complaint as the only text on record.

Then look at what assistants actually say about you. Discoverable's free visibility check runs your brand through Claude with live web search, no login required, and shows how you are currently described. Paid plans re-check automatically every week and keep a score trend history, so you can see whether a review push moves your score.

Incentivised and fabricated reviews are against every major platform's terms, and in the biggest markets they are now against the law. The FTC's rule banning fake reviews took effect in the United States in October 2024 with civil penalties per violation, and the UK's Digital Markets, Competition and Consumers Act made fake and undisclosed incentivised reviews illegal from April 2025.

For AI visibility the deeper problem is persistence. Text scraped into training data survives deletion: a model trained on a batch of purchased five-star reviews, or on coverage of your brand being caught buying them, carries that text for as long as the model runs, and no takedown request reaches it. Getting flagged also becomes content in its own right. 'This company was flagged for fake reviews' is exactly the sentence an assistant reaches for when someone asks whether you can be trusted.

The review strategy that survives the AI era is the slow one: real customers writing in their own words on the surfaces where buying questions get answered. It compounds, and unlike a bought rating, nothing about it needs to be hidden from a crawler that reads everything.