How to measure whether AI recommendations are working.

Start from the answers.

In this article, AI recommendations means ChatGPT, Gemini, Perplexity or Google's AI Overviews naming your business when a buyer asks what to use. It does not mean the output of a product recommendation engine on your own site, which is the older sense of the phrase.

AI assistants publish no impression counts, and Google folds clicks from AI Overviews and AI Mode into the ordinary Web results in Search Console (Google Search Central). So measurement starts from the answers themselves. Ask the same buyer questions on a schedule, store the answers, and count what they say. The numbers below all come from that record, and how to track whether AI recommends your business covers how to build it.

Mention rate.

The share of answers that name your business. Track 15 questions daily for a week and you have 105 answers. Named in 21 of them is a 20% mention rate.

Write it with its sample, always: "21 of 105 answers", not "20%". Our own rate on four questions about AI visibility tools was 0 of 42 answers, from 17 to 26 September 2026. That number is the reason this blog exists.

Visibility rate by day.

The share of your tracked questions where you were named or cited on a given day. Each question counts once per day, so one talkative answer cannot inflate it, and it answers the question owners ask first: on how many of my buyers' questions did I show up today?

Read the latest day next to the average of the last seven. The latest day tells you where you stand; the average tells you whether the latest day was typical.

Citation share.

The share of answers that link to a page on your domain as a source. A citation means the assistant read your page while it built the answer.

Watch which one moves first. If citations rise before mentions do, the assistant is finding your pages and still naming a competitor, and the fix is on the page: make it answer the buyer's question outright.

Question coverage.

The share of tracked questions where you were named at least once in the window, such as 4 of 15 questions over 30 days. Coverage shows breadth. A high mention rate on two questions and nothing on the rest reads differently from a modest rate spread across all of them, and it calls for different pages.

Share of voice.

Your mentions as a share of all the businesses named in the same answers. In our tracking, 40 answers named businesses in this category 284 times, and Semrush accounted for 22 of them, 7.7%.

Share of voice holds steady when answers get longer or shorter, which a mention rate does not. If an assistant starts listing ten tools where it used to list five, your mention rate can rise while your share falls.

Position and prominence.

Whether you appear as one of the answer's picks or in passing, and where in the list. In our tracking, 273 of the 284 business mentions were recommendations: an entry in a list, a heading or bold text.

In a category like that, being on the list is the whole question. In a category where answers name brands in passing, record the difference, because a passing mention next to a competitor's recommendation counts for little with the buyer.

What the answer says about you.

Read the sentence that names you. An assistant can recommend you for the wrong reason: an old price, a feature you dropped, an audience you do not serve. A rising mention rate built on a wrong description sends you buyers who leave.

If the description is wrong, look at the pages the answer cited. The assistant took the claim from one of them, and the fix belongs on that page, on your own site or on a list that describes you.

Answers given without searching.

Some answers come from the model's training data, with no search at all. Nothing you publish this month changes those answers. Count them on their own and leave them out of the rates above, so a question the assistant never searches for does not drag down a number that no page can lift.

In our 42 answers, Gemini searched before all 42, so this group was empty. In another category or on another assistant it may not be, and a tool that folds these answers into your rate hides the one part you cannot fix.

Sample size decides what a rate means.

Two answers out of seven and twenty out of seventy are both 29%. The first swings to 14% or 43% on a single answer; the second moves to 27% or 30%.

So write each rate with its denominator, compare windows of equal length, and wait for two weeks of daily data before you call a change. A move from 2 of 7 to 3 of 7 is one answer. A move from 20 of 70 to 30 of 70 is a trend.

Compare against the names that did show up.

A rate on its own does not tell you whether 20% is good. Put it next to the businesses the same answers named.

In our tracking of AI visibility tools, the most-named business, Semrush, appeared in 22 of 40 answers. The tenth most-named appeared in 9. That spread is the range to aim for in this category, and your own category will have its own. The full table is in who Gemini names when you ask for an AI recommendation tracker.

Connect it to traffic, with care.

Three sources show whether recommendations turn into visits.

• Your analytics. Visits from chatgpt.com, perplexity.ai and gemini.google.com appear as referrals. They undercount: a reader who copies a name into a search box arrives as search traffic.
• Google Search Console. It includes clicks from AI Overviews and AI Mode in the Performance report under the Web search type, mixed with ordinary results.
• Searches for your name. A reader who copies your name out of an answer searches for it. A rise in searches for your brand in Search Console, with no campaign behind it, suggests the answers are sending people your way.

None of the three tells you which question the visitor asked. The tracked answers do, which is why they come first.

Set a baseline before you publish.

Collect two weeks of daily answers before a new page goes live. Without a baseline, a change after publishing cannot be told apart from the day-to-day movement the numbers already had.

Keep the question set fixed while you measure. Adding or retiring questions changes the denominator, and a rate over a different set of questions is a different number.

A weekly routine.

1. Read the week's numbers with their samples: mention rate, visibility rate, citation share and coverage.
2. Open the answers on the questions that moved, and read what changed in the text.
3. List the questions where a competitor was named from a page the answer cited. Those are the pages to write next.
4. Note what you published and when, so a change two weeks later can be lined up against it.

If writing those pages is the step that stalls, outrankfast's Grow plan drafts the page for each question a competitor wins, and your coding agent can read it over MCP.