Citation signals

How customer reviews feed AI recommendations

By Arnav Mukherjee, founder of TofuBofu · July 6, 2026

A common piece of AEO advice is "get more reviews and AI will recommend you." It sounds obviously true, so most founders nod and move on. When I looked at the actual data, it turned out to be half right in a way that changes what you should do. Reviews matter enormously, but not in the direction the advice implies. More reviews will not lift you up the list. Having no reviews will keep you off it entirely.

The cleanest way to hold this: reviews are a gate, not a lever. They decide whether you are eligible to be recommended at all, and then they mostly stop mattering for where you rank. Getting that distinction right saves you from pouring effort into review volume that does not move anything.

The gate: be listed, or be invisible

Quoleady analyzed the tools ChatGPT names when asked for alternatives to popular software. The presence numbers were near-total: 100 percent of the tools had Capterra reviews, 99 percent had G2 reviews, and 78.8 percent had a Wikipedia page. Read that the right way. If you sell software and you are not on G2 or Capterra, you are almost never in the answer. The review-site listing is functioning as a category membership card. No card, no consideration.

This is why "should I bother with G2" is the wrong question for most software companies. The listing is not a nice-to-have marketing channel, it is part of how AI engines confirm you are a real player in your space. For services firms the same logic applies with different platforms: Clutch, Google Business Profile, and often G2 are the membership cards that say "this is a real firm buyers actually use."

Not a lever: why more reviews will not lift you

Here is the part the standard advice gets wrong. Once you are through the gate, review metrics barely predict where you land. In the same analysis, G2 review count correlated with placement at -0.16 and Capterra review count at -0.21. Star scores were no better: G2 review score at -0.11 and Capterra score at +0.02, which is effectively zero. These are weak, and some are slightly negative. More reviews and a higher average did not push tools up the list.

Domain rating, by contrast, was the strongest signal in the same study, which fits the broader pattern that authority and mentions decide ranking while reviews decide eligibility. So the takeaway is not "reviews do not matter." It is "get enough real reviews to clear the gate and describe you well, then spend your marginal effort on the things that actually move placement."

What engines do read from reviews

Count and average are the wrong things to optimize. What an engine can actually use is the content of recent reviews: the specific outcomes, features, and situations customers describe. A review that says "cut our response time from four hours to under one" hands the model concrete, quotable language it can attach to you. Recency signals that you are still good now, not five years ago. Sentiment and the topics reviewers raise shape how an engine describes you. This is the real mechanism: reviews feed the vocabulary AI uses to talk about your firm, which is far more valuable than one more five-star with no words.

The line you do not cross

Because reviews gate inclusion, the temptation to manufacture them is real. Do not. One study found that more than a quarter of G2 reviews since ChatGPT launched were likely AI-generated, and platforms and regulators are responding by tightening authenticity checks. Fabricated or purchased reviews violate platform terms and consumer-protection law, and when they are caught the listing gets cleaned out, taking your gate pass with it. The entire reason reviews carry weight is that they are supposed to be real. Fake them and you are spending money to build a liability. Earn them and you build a durable asset.

What to do