Reviews decay without anybody touching them.

Do reviews change how AI systems describe a business?

They are public, structured, dated, attributed and written by people who are not the business. For a system trying to establish whether a company is real and competent, that combination is unusually strong evidence, and it is read alongside everything else rather than as a separate score.

What is worth tracking, beyond the average

• Recency. A strong average built entirely from reviews that are two years old describes a business that used to be good. This is the signal that decays without any event, and the only one where doing nothing is itself the change.
• Distribution over time. Twelve reviews in one week and none since reads differently from one or two a month for two years. The second pattern is far more useful evidence, and far harder to manufacture.
• Whether the business responds, and how. Responses are public statements by the company, written in its own voice, on a platform it does not control. They are read.
• What the text says, not just the stars. Reviews that name the service, the location and the problem solved are the ones that can be quoted in an answer about that category.
• Consistency with the rest of the entity. Reviews describing a service the website no longer mentions, or a location that closed, put the business back into the entity-consistency problem from a direction nobody watches.

The line that does not move

Encouraging satisfied customers to leave an honest review is legitimate and is part of the work at the levels that include it. Selecting who gets asked based on how happy they are, offering anything of value in exchange, or writing them is not, and the fact that it would probably work is not an argument. A business whose proposition is being trustworthy cannot buy its evidence of being trustworthy.

The same logic covers a request we decline regularly: asking for a review before the work is finished, or attaching the ask to an invoice in a way that makes it feel required.

What tracking changes here

The failure is not a bad review. It is a quiet stall, where the pipeline that produced reviews stopped and nobody noticed for two quarters because the average did not move. Averages hide stalls. A tracked recency reading does not, and the correction while it is one month old is a conversation rather than a campaign.

That is the pattern the whole of AI visibility tracking runs on: catch the slow signals early, because the slow ones are the ones nobody is notified about.

Which platforms count, and why the answer is not all of them

The platforms that matter are the ones a system trying to verify this business would actually consult: the dominant general platform in the market, the one specific to the industry where one exists, and any platform the business itself points at from its own site. A profile nobody links to and nobody reads is not evidence, and spreading thin coverage across nine platforms is weaker than depth on two.

There is a practical limit worth naming. What can be read depends on what each platform chooses to expose, and several forbid automated collection outright. Where that is the case we read what is published and dated, and where we cannot read at all the row says not observable. A tracker that reports a confident number for a platform it has no access to is estimating, and an estimate that arrives every month starts being treated as a measurement.