One story before the argument.
I got called in on this and I couldn’t fix it, which is why I’m still turning it over eight months later. Everything the company did was right. Good agency, correct strategy, clean execution. And the machine simply refused to accept it. That’s the part I want to talk about, because I think a lot of people are about to spend money solving the wrong half of it.
– Ayush
Late in 2025, got a call from a $40M ARR B2B SaaS company. Eight years old, solid product, real customers.
Their problem was a word. The market had quietly reclassified them as bossware. That’s the category of tool that screenshots employee laptops and reports back. Nobody wants to buy bossware. Nobody wants to admit they bought bossware.
So they did the expensive, sensible thing. Six figures to a good agency. Every page rewritten. The word “monitoring” scrubbed off the domain. A clean relaunch as a workforce wellness platform.
By Februrary, their organic pipeline had fallen off a cliff.
The CEO was certain it was a site-indexing bug. So I ran the only diagnostic that mattered, and I didn’t open their website to do it.
I opened an answer engine and asked what they were. It said employee monitoring software, used to track remote workers. I asked for the best workforce wellness platforms. They weren’t on the list.
Not ranked low. Not there.
Now, the obvious read on that story is that they had an AI visibility problem. And that read is correct, which is the annoying part, because it’s also the read that will cost them another six figures.
What they were missing was prompt availability. Prompt availability is the odds your brand is the name in the instruction, rather than a candidate in the evaluation. It is not the same thing as being easy for a machine to find, and the gap between those two is what the rest of this is about.
What is AI availability, and who is saying it.
There’s a term for what they were missing. It’s AI availability, and it turned up around late 2025.
Search Engine Land put it plainly: after mental availability and physical availability, there’s now a third kind. The likelihood that an AI system recommends your brand when somebody is ready to buy.
Toluna’s definition is tighter. The probability that an AI system can correctly identify, retrieve, understand, compare and recommend a brand in response to a relevant prompt.
I have no argument with either. Both are describing something real.
And the supporting evidence is better than most things in this space. Toluna cites work putting roughly two-thirds of brand visibility inside language models down to long-term brand equity. That lines up almost exactly with the old IPA finding that around 60% of marketing effectiveness is long-term. (I haven’t opened the underlying study, so treat that pair as a strong lead rather than a fact.)
Search Engine Land makes an even better point. Category entry points, the moments of need that trigger a buyer to think of a category at all, are now expressed as prompts rather than as keywords. “Where can I find sustainable running shoes for flat feet” is not a keyword. It’s a buying situation, typed.
That’s a good observation and I wish I’d made it.
So the consensus holds. Being retrievable, interpretable and recommendable inside these systems matters. Everything above is correct.
Now let’s follow it one step further than the people selling it do.
What has to happen before a machine can recommend you.
Read Toluna’s definition again and watch the verbs. Identify. Retrieve. Understand. Compare. Recommend.
There’s a comparison in there.
Which means being recommended is not a state. It’s an outcome. It’s the outcome of an event that had to run first. A machine lined you up against everything else in your category and picked.
So when you optimise for AI availability, here is what you’re buying. You’re buying a better position in a comparison. Not an exemption from it.
That distinction sounds small. It isn’t, and the economics are where you see it.
Why the comparison is the part you lose.
Picture the two versions of the same request, from the same person, on the same morning.
“Find me a project management tool under fifty dollars a seat.”
“Set up Linear for my team.”
Same need. Same category entry point firing in the same head. Two completely different outcomes.
In the first, the agent goes off to evaluate the field on price and specification. That’s a comparison you mostly lose. There’s almost always something cheaper. Something with one more integration. Something that reads better on a feature grid. And the machine is very good at finding it.
Worse, you lose it again tomorrow. And on the next query. And on every query after that, against every competitor running the same optimisation.
In the second, no comparison ran. The human named the tool. The agent isn’t choosing, it’s executing. Every dollar of evaluation and every ruthless price check simply never happened.
That’s the whole distance, and it’s what prompt availability measures. Notice it has nothing to do with how well you optimised. It’s a difference in what the person typed.
What the comparison fight looks like up close.
I want to show you the thing you’re signing up for, because it’s worse than most people assume.
Wharton’s Generative AI Labs ran a test on how susceptible frontier models are to persuasion. 126,000 conversations. Claude Haiku 4.5, GPT-5 mini and Gemini 3 Flash.
The test is sharper than a marketing example would have been, because they weren’t studying product recommendations at all. They tried to talk each model into complying with requests it is trained to refuse. Including handing over the synthesis route for a scheduled drug.
Adding one classic persuasion principle to the request raised compliance from 35.3% to 51.3%. The authors call the models parahuman. They respond to social influence the way people do, because they learned our patterns from a century of our own text.
Be careful what you carry out of that. It’s a safety context. It doesn’t transfer cleanly to a purchase.
A drug-synthesis refusal is a boundary the model is built to hold. A product recommendation has no such guardrail. So the lift there could be larger, smaller, or shaped nothing like it.
The transferable finding is narrower and more durable. The reflex is there at all. Reasoning didn’t remove it, and it held across three different vendors.
Which tells you the comparison layer is a knife fight. Against opponents running your playbook, with the vendors patching as they go. You can win it on any given day. You will never be allowed to stop.
Prompt availability means you were never in that fight.
And I know how this looks. I’ve spent a piece complaining that AEO, GEO and LLMO are three names for one practice, and here I am adding a fourth availability. So let me be precise about the difference. Those three renamed a job everybody was already doing. This names a distinction nobody had drawn, between being compared favourably and not being compared at all. If the distinction turns out to be wrong, drop the term with it.
The strongest objection to prompt availability.
Here’s where I have to give some of that back, because there’s a hole and it’s a good one.
The names people type are increasingly the names the machine taught them.
Ask an agent about project tools twice. The third time you may well type “set up Linear.” And you’ll type it because the agent put Linear in front of you the first two times.
If that holds, prompt availability isn’t upstream of the comparison at all. It’s downstream. It’s just the delayed output of the fight you thought you’d skipped.
I think both are true, and which one you get depends entirely on where the name was learned.
A name earned through a channel the agent never mediated is upstream. A product somebody used. A recommendation from a person they trust. A reputation built before the query existed.
That name is hard for a machine to recalculate. It never calculated it in the first place.
A name the agent supplied last Tuesday is not prompt availability. It’s an echo, and it belongs to whoever won the comparison that produced it. (Which is most of them, if we’re honest.)
Why the rebrand failed, and what would have worked.
They changed the reference document. They didn’t change a single category entry point.
Every cue that fired in a buyer’s head still pointed at the old name. So did every source the machine consulted when one of those cues got typed. [Eight years]{.mark} of reviews, forum threads and backlinks against [fifty new pages]{.mark} of their own copy.
And fifty pages of new copy is a claim about yourself. Which is the one input the system weights lowest.
Here’s what I’d have told them, if the answer had been cheaper. They had an AI availability problem and a prompt availability problem, and they spent the money on the first one. Optimising AI availability would have improved their position in a comparison they were always going to lose. The machine had them in the wrong category. No amount of retrievability fixes being retrieved as the wrong thing.
The expensive, slow, unglamorous fix was to go and change what the rest of the internet said about them. That’s brand work. It takes years. And it never shows up on a dashboard.
So I’d stop asking how to get recommended. Being recommended means the comparison ran, and the comparison is the part you keep losing.
Ask what it would take to build prompt availability instead. To be the name somebody types before any of it starts.
If you work out how to do that faster than I have, come and tell me.
