Before optimising for anything, it helps to have a working model of what is happening. Assistants are not consulting a ranking table. They are assembling an answer, and understanding the steps tells you where you can intervene.
The four steps
When someone asks an assistant who provides a treatment near them, roughly this happens:
- Interpretation. The question is turned into an intent and a location. Vague questions get interpreted generously, which is why answers vary.
- Retrieval. The assistant gathers material: search results, map data, directory records, review content, and its own training data.
- Resolution. It works out which named entities exist and which retrieved facts belong to which entity. This is the step that decides your fate.
- Composition. It writes a short answer naming a small number of options, usually with a sentence of justification each.
Why resolution is where you win or lose
Step three is the one nobody optimises for and the one that matters most. The assistant has to decide that a scatter of records describes one real practice.
If your name, address, and phone agree across those records, resolution is easy and you become a confident candidate. If they disagree, the assistant faces a choice between asserting something it is unsure about and naming a practice it can verify. It reliably prefers the second.
This is the crucial asymmetry with traditional search. A results page can hedge by listing ten links and letting you sort it out. A three-name answer cannot hedge. Ambiguity does not push you down the list, it removes you from it.
Why answers are inconsistent
Ask the same question twice and you may get different practices. This is normal and has real causes:
- Generation is probabilistic, so phrasing and randomness shift the output.
- Location inference differs between sessions and devices.
- Live retrieval returns different sources at different times.
- Training data and retrieval indexes update.
- Personalisation and conversation history influence the answer.
The practical consequence, developed further in the measurement lesson, is that a single prompt tells you almost nothing. You need repeated prompts across engines to see a real pattern rather than a lucky draw.
What actually shifts the answer
In rough order of influence:
- Entity consistency. The dominant factor, because it governs resolution.
- Presence in the sources being retrieved. Map profiles, the significant directories, and your own site. You cannot be assembled from material that does not exist.
- Explicit, factual statements on your website. Assistants extract facts. A page plainly stating which treatments you provide, where, and for whom is easy to use. A page of atmospheric brand language is not.
- Structured data. Removes ambiguity about what and where you are.
- Reviews. Used both as prominence evidence and as material for the justifying sentence.
- Corroboration. The same facts appearing across several independent sources.
What does not work
- There is nothing to submit. No listing, no feed, no paid inclusion.
- Keyword stuffing. These systems read meaning, not density.
- Writing copy addressed to the AI. Text asking to be recommended is worthless and looks manipulative.
- Volume of thin content. Thirty vague pages lose to three specific ones.
- Anyone selling guaranteed placement. Nobody can guarantee it, because there is no placement to buy.
The unglamorous conclusion
The best preparation for AI visibility is doing chapters two through four properly. Accurate profiles, consistent data, real coverage on the sources that get retrieved, and a website that states facts plainly. There is no separate AI channel to buy, which is genuinely good news: the work compounds with everything else.
The next lessons look at each engine, since their source weighting differs.