AI Discovery Is a Memory Problem, Not a Search Problem

Ask ChatGPT for the best agency for your problem, close the window, and ask again tomorrow. Around 30% of the brands named the first time survive to the second identical query. If the model were consulting a ranking, that number would embarrass it. It is not consulting anything. It is building the answer live, from cues, and the discipline that governs which cues exist has a fifty-year-old name in marketing science: mental availability.

The industry looked at AI answers and reached for the nearest familiar frame: this is search, so we need optimization, so we need Answer Engine Optimization. The frame is wrong at the first word. Search engines indexed documents and retrieved them. Language models store no document, no list, no verdict about you. Treating an improvising system as an index is why so much AEO advice reads as SEO with the nouns swapped.

The model is not looking anything up

There is no table inside ChatGPT with your brand's score in it. When someone asks for a recommendation, the model generates the most plausible continuation of that question, assembled from precedent in its training data and whatever third-party text it retrieves in the moment. The answer did not exist before the question. It will not exist after it.

The instability is the proof. Recommendation-stability studies in 2026 found that roughly 30% of cited brands persist to the next identical query, and about 20% across five runs. An index would return the same result five times. An improviser returns five improvisations. There is no ranking to win. There is a memory to occupy, so that when the improvisation happens, your brand is the material it reaches for.

Buyers never looked anything up either

Here is the part the dashboard vendors miss: the buyer works the same way. The behavioural scientist Nick Chater made the argument at book length in The Mind Is Flat (2018): buyers do not carry a warehouse of settled opinions waiting to be surveyed. Judgments are constructed at the moment of asking, out of stored traces and whatever cues the question supplies. Ask a buyer what they think of your brand and you are not measuring an opinion. You are watching one get manufactured.

Chater's claim needs one correction, and the correction is the useful part. Memory research shows that plenty is stored. What is stored is traces and cues, not verdicts. The trace is durable; the judgment is assembled from traces at the moment of asking. This resolves the fight between Chater and the memory scientists, and it tells a marketer exactly what is and is not in their control. You can supply traces. You cannot install a verdict, in a buyer or in a model.

So the two systems that decide your commercial fate, the buyer's head and the model's weights, share one architecture: cue-driven construction. One set of laws governs both.

The laws were published between 1973 and 2018

Tulving and Thomson¹ compressed the first law into one sentence in Psychological Review, and it transfers to machine answers without editing its logic:

"What is stored determines what retrieval cues are effective in providing access to what is stored." — Endel Tulving & Donald Thomson, Psychological Review, 1973

Replace "stored" with "written about you on the surfaces a model reads" and the sentence describes AI search visibility. Collins and Loftus² supplied the second law, and it is not a metaphor for brand associations. It is the model those associations were named after.

This is why the memory frame outperforms the search frame in practice. Search thinking asks "what query do we rank for?" and produces keyword pages. Memory thinking asks "which buying situations is this brand encoded in, and what is it adjacent to?" and produces a different work order entirely: get encoded into the situations, on the surfaces the improviser reads.

Why the dashboards cannot tell you this

The AI-visibility tool market measures citations: were you mentioned, at what rate, against whom. Useful, and we use those numbers too. But a citation count is the symptom. The cause is the cue supply, and the cue supply lives almost entirely off your website: about 85% of what AI assistants say about brands comes from third-party sources³. Reviews, communities, press, comparison pages. The model forms its improvisation from what third parties write about you, in the contexts where they write it.

In the Brand Distinctiveness Framework, that split has names. The Entity layer makes a brand resolvable, Topic Authority makes it trusted, and Voice makes it recognisable without the logo. AI search visibility is the outcome those layers produce, never a lever you pull directly.

Counting citations without engineering cues is watching the scoreboard without playing. We wrote about the difference between being cited and being recommended in Answer Engine Optimization: How to Get Recommended by AI, and about the entity side of the cue supply in Why Your Brand Isn't Showing Up in ChatGPT. The pattern across both: measurement is an engineering problem, but knowing which cues to build is marketing judgment, and the judgment came first by fifty years.

The cue audit: where to start

Run this before buying any tool. It takes an afternoon.

1. Write the ten situations that trigger your category. Not keywords. Situations: "our pipeline died", "the board asked why we're invisible in ChatGPT", "the audit found duplicated content". These are your category entry points, and each one is a retrieval cue.
2. Ask each situation to the engines. Put the situation, phrased as a buyer would phrase it, to ChatGPT, Gemini and Perplexity. Record who gets named. That roster is the memory you are competing with.
3. Check whether your brand was ever encoded in each situation. Search your third-party footprint for each entry point. Most brands discover they have one key and a lot of doors: strong association with their category name, nothing linking them to the moments that start a purchase.
4. Map your adjacency. List what your brand sits next to in public text: which problems, which comparisons, which named alternatives. If the answer is "our own product terms", your network has one strand.
5. Repeat the engine pass on a schedule. One run tells you nothing durable, because the answer is improvised each time. The trajectory across runs is the real reading.

The gap between steps 1 and 3 is the work. Closing it means getting encoded, situation by situation, on the surfaces models read. That is slower than publishing another blog post and it is the only path that compounds.