Two prompts, typed by the same finance lead, three weeks apart.
The first: best expense management tools for a 200 person company. The second: is your product any good for a company our size.
The first prompt decided if you would ever be considered. The second arrived after that decision was made, and its answer was close to guaranteed the moment your name went into the box. Most AI visibility reports add both into one percentage, report it monthly, and the budget follows whichever half moves easiest. That is the second prompt.
I have been writing software for 30 years. Every metric I have watched go wrong went wrong the same way: somebody let the easy part of the denominator grow.
The thesis
Branded prompt visibility is not an asset you build. It is a receipt for a decision made earlier, in a prompt that did not contain your name. A blended AI visibility score counts the receipts as wins, so most programmes are optimising and reporting the half of the prompt list they had already won. Split the score in two and put almost all of the budget on the unbranded half.
The mechanism
A buyer types your name for one reason: you are already on the list. 6sense surveyed about 4,000 B2B buyers for its 2025 Buyer Experience Study. Buying groups had filled four of their five shortlist spots on day one, first contacted a seller 61% of the way through the process, and bought from the vendor they ranked first 77% of the time. The list forms before anyone is talking to you. In 2026 a growing share of it forms in an answer box.
Now look at what the engine has to do in each case. When the prompt names you, retrieval is trivial. The engine searches your name, fetches your own pages, and talks about you, because you are the subject of the question. When the prompt names a category, the engine has to find you first, and it finds you only if something in its retrieval path points at you.
A paper Benjamin Tannenbaum posted to arXiv on 19 September 2026 measured exactly this. He modelled 34,960 unbranded prompt observations across 75 monitored projects, run repeatedly on GPT and Gemini between June and September. When neither the brand nor its domain appeared anywhere in the live retrieval path, the brand was mentioned 2.8% of the time on GPT and 3.8% on Gemini. With its own domain cited, 49.0% and 58.4%. With its own domain cited and the engine running a follow up search on the brand name, 91.4% and 100%.
The second finding matters more. Visibility persists. A brand that was absent in the previous run and had no current exposure was mentioned in the next run 1.6% of the time on GPT. A brand that was present before and exposed now: 80.5%. The unbranded answer has incumbents, and absence compounds.
The arithmetic that hides it
Take a tracked list of 100 prompts. This is a hypothetical, but a common shape: 40 name the brand, 60 do not. The brand appears in 90% of the branded answers and 5% of the unbranded ones. Blended visibility is 39%.
Now improve either half. Lift branded from 90% to 98% and the blend goes to 42.2%. Double unbranded from 5% to 10% and the blend goes to 42%. The board pack shows the same 3 point gain both times. One change added nobody to a single shortlist. The other doubled the number of buyers who could find you without already knowing your name.
The branded half is also cheaper to move. Fix a wrong price on your own site, publish a comparison page, and the branded number jumps inside a fortnight. Getting into a category answer means getting into pages you do not own. A team paid on the blend will pick the first job every time, and on the incentives it was given, it is right to.
Three data sets, one shape
The NeuroRank AI Visibility Index 2026, reported on 6 October, tested 122 brands across nine sectors on ChatGPT, Gemini, Claude and Perplexity, using 7,580 prompts. Results where the brand was named in the prompt reached high inclusion 56.2% of the time. Results without the name: 17.8%. Of 82 brands assessed both ways, 51 cleared 50% when named. Three cleared it unnamed. No sector reached 30% on unnamed prompts. The authors say this is not a controlled test of adding a name. It does not need to be. It is the gap a blended number averages away.
BuzzStream’s mentions and citations study, published in June, scored 12,000 responses across Google AI Mode, AI Overviews, Gemini and GPT-5.4-mini for about 200 brands. On single brand prompts such as “Is Asana good”, a brand mention came with a citation 39.0% of the time. On list and category prompts such as “best project management tools”, 7.2%. Even when a brand’s own page was cited in a category answer, the brand was named in the text only 36.5% of the time, against 79.1% when the prompt was about the brand. In the category answer, being a source is not the same as being recommended.
Tannenbaum’s persistence result is the third. Read together, the three say one thing. The branded half of the list is close to saturated for anyone with a working website. The unbranded half is where brands are absent, where absence sticks, and where most of them are losing.
My rule follows from that. When an AI visibility number reaches a board pack, I want the prompt list before I read the number, and I want every prompt tagged by whether a brand name is in it.
Where this breaks
Branded answers are not free. “Is it any good” is the last prompt before a buyer books a demo, and an engine that quotes last year’s pricing or a dead integration loses the deal at the final step. That is real work. It is also a correctness job, not a visibility job, and it should be priced as maintenance: a fixed quarterly line to find and fix what the engines get wrong about you, not a growth budget.
The thesis also weakens for category leaders. If most of your buyers arrive already knowing your name, the branded half carries more of the demand and the split matters less. Most companies paying for an AI visibility programme are not in that position, which is usually why they are paying.
And prompt lists are synthetic. Nobody outside the engines knows real prompt volumes, so the 40 and 60 in my example are a choice. That is the point. Somebody chose the mix in your report too, and the mix sets the number.
What I would do on Monday
Export the tracked prompt list from whatever tool you use and tag each prompt three ways: names your brand, names a competitor, names nobody.
Report two numbers from now on. Branded accuracy: is what the engine says about you correct. Unbranded inclusion: do you appear when nobody asked for you. Retire the blend.
If fewer than half the prompts name nobody, rebuild the list from sales call notes, using the first question buyers say they asked before they knew any vendor.
Give the branded half a fixed maintenance budget. Point everything else at the pages already sitting in the retrieval path for your category prompts.
Track unbranded inclusion run over run, per engine. Absence persists, so the first movement is the one that counts.
Two prompts, three weeks apart. The second one was always going to mention you, because the buyer typed your name. The first one decided whether the buyer ever would.
Score the first one.
If your AI visibility report still carries one number and nobody has split the prompt list behind it, pick a slot and bring the export.