How should an agency report AI visibility to clients?
All guidesUpdated 6 steps
TL;DR
Report it on a locked set of prompts, with the method named and every number dated. Say whether each number comes from search results, which is a proxy, or from asking the models directly. Show failed checks as not measured, never as zero. Split branded and unbranded prompts, tie each change to work you shipped, and never call a search-result appearance a citation.
Clients ask a simple question: does ChatGPT recommend us? The honest answer is rarely simple. Answers vary between runs, accounts and locations, and no console reports AI mentions the way Search Console reports Google clicks.
That makes AI visibility easy to oversell. A report that looks precise but changes its prompts every month, or fills gaps with zeros, will be caught out sooner or later. A report that states its method plainly can be trusted when the number goes down as well as up.
Step by step
- Step 1
Agree a locked prompt set with the client
Write about ten prompts with the client: half branded-adjacent, such as comparisons with a named competitor, and half unbranded, such as the best tool for a specific audience. Get the client to sign off, then lock the set. Changing even one prompt starts a new baseline, and the report should say so.
- Step 2
Name the method on every number
If you check whether the client's domain appears in the search results for each prompt, call it search visibility on the locked set, a results-page proxy. If you ask the models directly, say which models and when. Never blend the two into one score.
- Step 3
Count only the checks that ran
Write the score as a count of the prompts that actually ran, in the form "6 of the 10 that ran". If some checks failed, the denominator shrinks, and the report names the failed prompts and the reason. A missing check is not measured, and reporting it as zero tells the client something false.
- Step 4
Split branded and unbranded
Show the two halves separately. Branded-adjacent prompts mostly confirm that people who already know the name can find it. Unbranded prompts show whether strangers hear about the client, and a total that rises only on the branded half is not the growth it looks like.
- Step 5
Show who took the answer slots
List the domains that appeared where the client did not, with the number of prompts each appeared in. Many will be directories, review sites, forums or publishers rather than business competitors. Those may be places to get listed, not rivals to beat.
- Step 6
Tie every movement to work shipped
Next to the score, list what changed on the site since the last report: pages improved, crawler access fixed, listings corrected. When a number moves with no work behind it, say the cause is unknown. RankingTop's paid plans run ten locked prompts on a schedule and keep a dated readout of every run.
Questions people ask next
What is the difference between proxy and direct measurement?
A proxy checks whether the client's pages appear in the search results for a prompt. It matters because an assistant that searches can only quote pages it can retrieve, but appearing there does not prove an assistant quoted them. Direct measurement asks the models the prompt and reads the answer, which is closer to the real thing and still varies between runs.
Can I report AI Overviews traffic from Search Console?
Not as its own number. Google's documentation has described AI Overviews and AI Mode appearances as counted within the regular web search totals in Search Console, not as a separate report. Report Search Console as Google search overall, and keep AI assistant checks in their own section.
What do I do when the number drops?
Report it with the same method and the same prompts, and say what you know about why. First check that the drop is real: failed checks, a changed prompt set or a new model version can each move the number. A client who sees honest drops will believe your gains.
Should I promise a client they will be cited by ChatGPT?
No. Nobody outside OpenAI controls which sources ChatGPT names, so that promise cannot be kept honestly. Promise the work instead: crawler access fixed, answer-first pages shipped, listings corrected, and a fixed measurement that shows what changed.