AI visibility for agencies

Measuring one brand and measuring twenty are not the same problem. Here is what changes, and what we learned running a portfolio.

One account groups several domains behind a single login. The plan bills the account, not each domain, and every client sees only their own.

What changes when you measure a portfolio

One example from August 2026, measured on our own portfolio: Google AI Mode went from citing 12.22 sources per answer to 4.08 in a single week. Eleven brands across ten sectors lost a third of their visibility at once, having done nothing. With one client that is a scare. With twenty it is a whole morning explaining drops that are nobody's fault — unless your tool tells you the change came from the engine.

That is why every measurement carries a stamp of how it was taken (which engine, which transport, how many repetitions), and the report warns you when a series mixes methods and has stopped being comparable with itself.

One account, several domains

Each domain is measured with its own questions, its own competitors and its own history. What they share is the login and the plan.

What your client sees, and what you see

That isolation is not an interface preference: it is enforced on the server on every request. A client link asking for another domain gets an error, not an empty screen.

The panel with your branding

It is what lets the measurement be part of your service instead of looking like a third-party tool you forwarded.

The data, inside your reports

• CSV export by URL, to pull into Looker Studio or a spreadsheet with one formula.
• MCP server: ask your own AI about a client's data without opening the panel.
• Weekly email report with what changed and why.

Fixing a client's competitors without measuring again

It works because every answer is stored in full: re-scoring it against a different competitor list means re-reading text already on disk, not asking the engine again.

This matters more than it sounds. Nobody knows a client's competitors better than the person running the account, and that list is what the whole report hangs on: who shows up instead of them, share of mentions, the sector ranking. With one real client, of fifteen carefully reasoned competitors at onboarding the AI named only two; the ones actually filling the answers were five that were not on the list. Fixing it cost nothing and changed half the report.

What editing competitors does not change is the client's own visibility: it would be inexplicable for their percentage to drop because you added a rival.

How to justify it to the end client

The evidence on whether AI traffic converts better or worse than organic is contradictory, ranging from 13% worse to 31% better: in practice, parity. And the two ends of that range do not carry the same weight — the 13% worse comes from a peer-reviewed study of 973 sites and 164.8 million transactions (Kaiser and Schulze, Marketing Science, 2026); the 31% better, from a vendor analysis of 94 sites. Worth knowing before someone quotes it at you: anyone selling a number will always pick the second one. What is measurable is that the channel is still a tiny share of traffic and growing fast, and that positions are being handed out right now.

The argument that works in a meeting is not a projection: it is showing the client the real questions in their sector, who appears in them, and which sources the AI is reading to answer. That needs no justification.

And when the number drops without anyone doing anything

It will happen, and it is the most uncomfortable conversation there is. In the week of 22 August 2026, Google AI Mode citations fell 67% at once across all eleven clients we were measuring, simultaneously. None of them had touched their website. The engine had started giving shorter answers with fewer links.

What saves that meeting is not an improvised explanation: it is having the figure for the whole portfolio. One client dropping is that client's problem; eleven clients dropping in the same week on the same engine is the platform, and you can demonstrate that in two minutes if you measure all of them. Here is that case with its sample and its method.

Hence the boring part, and the one that actually protects an account: if the report you hand your client carries its margin of error, a drop inside that margin is a thirty-second conversation. Without a margin, it is a meeting.

Build your own or buy one?

We say that from experience: this engine was built exactly that way, and the expensive work was never firing questions at a model. It was everything else — noticing that an engine changed behaviour, telling a client's drop apart from an engine's drop, keeping series comparable across a change of method, and not publishing a number that does not hold up.

If you are going to build it, do it knowing where the cost is. And if you buy it, demand this of whoever you buy from: that they publish how they measure.

Frequently asked questions