Guide · Measurement
Understand AI visibility numbers
Understand mention rate, citation rate, share of voice and coverage using the answers and counting rules behind each figure.
Begin with the saved answers
Define the review window, question set and collection surface first. Keep a route back to every answer included in the total.
A chart without its underlying records can't show what was said. Start with the evidence before you interpret a change in the figure.
Write the review's scope at the top of your notes. Name the assistants, the country, the prompt set version and the first and last capture dates. A reader who sees only the figure can then tell what it covers.
Make sure each saved answer opens in full, with its date and assistant. An answer you can't reopen can't be rechecked, so leave it out of the count and say why.
Agree who holds the saved answers and where they are kept. A figure is only as good as the records someone can produce when asked.
Adding easier prompts can shift a rate without changing any old answer. Keep the previous version of the set in your record, and show what changed in the inputs.
When you have to change the prompt set, run the old and new sets side by side for one period. That overlap shows how much of any movement comes from the new prompts. Then mark the break in your records.
Changes outside your prompt set count too. If an assistant changes its product or your capture method changes, note the date. A shift on that date may come from the change rather than from anything about your business.
Define a mention
State which business name matches count as mentions and how repeated names are handled. Keep readable answers with no mention in the denominator.
Save uncertain matches for review and don't count them as definite. Keep the rule beside the count so another reader can follow it.
List the spellings your rule accepts. Include your trading name, any former name, common misspellings and the product names customers know you by. Decide whether naming the owner counts as naming the business.
Watch for a different business with a similar name. Check the town, the service or the web address before you count it. A mention can also be unfavourable, so record what the answer said about you in a separate column from the count.
Define a citation
A citation links to a source. Decide whether you count answers containing a link or individual links, because they give different totals.
Check the linked page belongs to the business when that's the question. A mention without a link is still a mention, so keep the two in separate columns.
Decide which addresses count as your own site. Subdomains, an old domain that now redirects and a booking page hosted elsewhere each need a ruling. A link to your listing on a directory is a citation of that directory, not of your website.
Open a sample of the links. A link can point to a page that no longer exists or to a page that never mentions you. Count it under your rule, then note the problem for the review.
Count places in ordered lists with care
Count explicit ordered-list positions under a stated rule. An answer with several rankings and an answer count are not the same unit.
Average place should use only the lists that included you. Say how many lists that was. An average from two lists and one from forty look the same on a chart.
Keep place and presence apart. Being named fourth in many lists and named first in one list are different findings. Report how often you were named first as its own count.
Show the denominator
Keep the count and total together. A missing or unreadable answer is a collection gap, not a completed answer that left the business out.
The denominator describes the group you observed. It doesn't represent every question real customers ask or every answer they might get.
Keep the denominator honest when you filter. If you look only at prompts that name a town, the denominator becomes the readable answers to those prompts. Write the filter beside the figure so nobody compares it with the unfiltered rate.
Keep the counts in the same row as the rate in any table you share. A rate copied without its counts is the first thing to lose its meaning when a figure is passed on.
Report prompts that name you separately
An answer to a prompt that already contains your name has an obvious reason to repeat it. Mixing those prompts with general ones lifts the mention rate without telling you anything about discovery.
Label each prompt as branded or unbranded and report the two rates apart. The unbranded rate is the one that describes whether you are named while someone is still choosing.
Branded prompts still matter. They show how an answer describes your business when a customer already knows your name. Read those answers for accuracy as well as counting them.
Show capture coverage
Show how much of the planned collection produced readable evidence. A rate can look stable while its underlying coverage has changed.
The other two stay missing. Don't treat them as answers without mentions. A missing capture may just mean a blocked request, not anything about how the assistant describes you.
List why each capture failed where you know the reason. A timeout, a blocked request and an empty response are different problems. If coverage is low, consider reporting the counts alone and holding back the rate until coverage recovers.
Keep the failed attempts in the record rather than deleting them. A later reader needs them to work out the coverage again.
Separate percentage points
Use the right unit when comparing percentages and keep the underlying counts. Do not describe an answer metric as a measurement of traffic or sales.
Include the counts for each period. Small samples swing sharply when a single answer differs, and a precise calculation doesn't make a sample representative.
Here is an illustration. Two mentions in ten readable answers became three in ten. That is a rise of ten percentage points, and a relative rise of fifty percent. Both statements are accurate, but the second sounds much larger. One extra answer produced the whole change.
When the counts are small, say so in the same sentence that gives the change. Readers can then judge the movement for themselves.
Read repeated checks over time
One answer on one day is a single observation. The same prompt asked on several days shows whether a mention holds or comes and goes. Record each day as its own row.
Decide in advance how many days of checks a prompt needs before you describe it as steady. Apply the same threshold to each prompt and each competitor.
Look at movement in both directions. A mention that comes back on alternate days is a different finding from one that stopped a week ago. The daily rows show which you have.
Where Google tools count AI visits
Visits sit outside your answer records, and the main tools file them in different places. Know where each one puts AI traffic before you compare figures.
Google says visits from AI Overviews are included in overall search traffic in Search Console. They are reported in the Performance report under the Web search type.
Google Analytics has a default channel called AI Assistant. Google describes it as visits from sources such as ChatGPT, Gemini, Deepseek, Copilot or Grok. It excludes Google's AI Overviews.
So a visit from an AI Overview and a visit from a chat assistant sit in different reports. Keep them as separate rows, and never add either one to a mention count.
Google's newer guide also documents a Generative AI performance report in Search Console. Check which report your property provides and its coverage.
Don't turn an answer count into a revenue claim. Use the figure to choose what to review, then check the statements and links underneath.
Give each measure its own source. Answers come from your saved captures. Visits come from your analytics or server logs. Enquiries come from your own forms, phone records or booking system.
Where you do compare answers and visits, put them side by side with their own dates and sources. A rise in both over the same weeks is two observations, not one result.
Write each figure so it can be checked
A figure that travels into a report or a meeting loses its context fast. Write it as a sentence that carries the context with it.
An example reads like this. Fern Workshop was named in three of eight readable answers, from ten attempts, on one assistant over one week. The counting rule was version two.
Anyone reading that sentence can see the denominator, the coverage and the scope. They can also ask for the eight answers and check each one.
If a figure can't be written this way, because a count or date is missing, say what is missing. A gap stated plainly is more useful than a number nobody can trace.
A worked example
Fern Workshop requested ten captures. Eight produced readable answers and three of those named the business.
| Detail | Observation | Review or next step |
|---|---|---|
| Mention rate | Three mentions in eight readable answers is thirty-seven point five percent. | Use the readable answer denominator and keep the counts. |
| Coverage | Eight readable answers in ten attempts is eighty percent. | Keep the two missing attempts apart from absence. |
The mention rate uses readable answers and coverage uses attempted captures, so the two rates answer different questions.
Common questions
What should zero readable answers show?
The rate is unavailable. With no readable answers there's nothing to calculate it from.
Can a higher rate show I have more customers?
No. Customer visits and business outcomes need their own evidence.
What should zero readable answers show?
The rate is unavailable. With no readable answers there's nothing to calculate it from.
Can a higher rate show I have more customers?
No. Customer visits and business outcomes need their own evidence.
Check your understanding
Three mentions occur in eight readable answers from ten attempts. Which mention rate fits?
Read the answer and explanation
- Thirty percent, treating missing answers as absence.
- Thirty-seven point five percent, with eighty percent coverage shown separately.
- One hundred percent because all three mentions were found.
Answer. Thirty-seven point five percent, with eighty percent coverage shown separately.
The mention denominator is eight readable answers, and coverage uses the ten attempts.
Answer and explanation
Thirty-seven point five percent, with eighty percent coverage shown separately.
The mention denominator is eight readable answers, and coverage uses the ten attempts.
Your next steps
Sources and method
A Carqo review method. Provider statements link to the relevant documentation.
- Google guidance on AI features Link checked 4 November 2026
- Google Analytics channel definitions Link checked 4 November 2026
- Google, optimizing your website for generative AI features on Google Search Link checked 4 November 2026