GEO Has No Rank, No Impression, and No Referrer. That Is Why It Gets Billed as Brand.

Everyone selling GEO services will tell you it is the new SEO. The pitch is easy to make, because the work looks similar: publish good pages, structure them clearly, get cited. The part that does not transfer is the reporting.

SEO reporting rests on three numbers that an AI answer does not produce.

The three numbers that go missing

Rank. A search result has a position. Position three on Tuesday and position seven on Friday is a fact you can plot, and everyone plotting it sees the same thing. An AI answer has no position. Ask the same question twice and you can get two different answers citing two different sets of sources, because the model samples rather than looks up. There is no row your page sits in. There is only whether it turned up in this particular generation, this time.

Impressions. Search console tells you how often you appeared for a query even when nobody clicked. That number is what makes zero-click search survivable as a discipline: you can still show the work happened. AI answers have no equivalent surface. Nobody reports how many times a model mentioned you, because nobody except the model provider is counting, and they are not sharing.

The referrer. This is the one people underestimate. When an AI answer names your product and the reader acts on it, the most common outcome is not a click on a citation. It is the reader typing your brand name into a browser a few minutes later, or remembering it next week. That arrives as direct traffic or branded search, with no thread back to the answer that caused it. Even when someone does click a citation, the referrer is inconsistent across hosts and often stripped.

Take away rank, impressions, and referrer, and you have removed the entire scaffolding that makes performance marketing performance marketing.

What you can actually count

Four things, none of which is traffic attribution.

Whether the crawlers read you. GPTBot, ClaudeBot, PerplexityBot and Google-Extended announce themselves in your server logs. Filtering your access logs for them tells you which of your pages are being fetched and how often. This is the cleanest signal in the whole exercise, and it measures the wrong thing: being read is not being cited. Treat it as a coverage check. If a page you care about has never been fetched, nothing downstream can happen.

Share of voice, by sampling. Write down the thirty questions a buyer would actually ask. Run them on a schedule against the engines you care about, and record whether you were mentioned and what was cited alongside you. This is a poll, not a census: the sample is small, the answers vary run to run, and you need repetition before a change means anything. What it gives you is a defensible trend line and, more usefully, a list of the competitors and sources the model reaches for instead of you.

Referral traffic from the hosts that send it. Some do. Segment it, watch it, and do not extrapolate from it. It is the visible tip of an effect whose size you do not know.

Branded search volume and direct traffic. The lagging proxy. If GEO is working, more people know the name, and that shows up here eventually, mixed in with everything else you did that quarter. It is real and it is uncontrolled.

Why that puts it in the brand budget

A performance line item survives a budget review by showing cost per outcome. GEO cannot produce that number honestly. What it can produce is presence: you are mentioned in this share of the questions that matter, alongside these competitors, up from last quarter.

That is exactly the shape of a brand argument. It is also how PR, sponsorships and category advertising have always been funded — on the reasoning that being known is worth paying for even when the ledger cannot show which pound bought which customer.

So GEO ends up filed there, and that has consequences worth naming. Brand budgets are approved on belief and cut first when the quarter is tight. They are reviewed annually rather than weekly. Nobody expects a brand line to defend itself with a dashboard, which means nobody builds one, which means the discipline stays unmeasured for longer than it needs to.

What to do about it

Set the expectation before the work starts, not at the first review. If you promise attributable traffic from GEO you will be arguing about the numbers within a quarter, and you will lose, because the numbers do not exist.

Build the sampling harness early. Thirty questions run monthly is a small amount of work and it is the only trend line you will have that responds to what you actually change. Start it before you make changes, so you have a baseline that is not reconstructed after the fact.

Keep the crawler logs. They are free, they are yours, and they are the one place where the raw fact of being read is recorded rather than inferred.

And write the pages anyway. The thing that gets a page cited by a model is close enough to the thing that got it ranked by a search engine — a clear answer to a real question, in one place, that a machine can lift without ambiguity. That work pays into both. It just will not send you a receipt.