AI Visibility
How AI Assistants Choose Sources and Which Businesses They Cite
An observation-vs-causality taxonomy and a repeatable prompt testing protocol to separate what really drives citations from noise.
Nobody outside the labs knows exactly how an assistant selects sources. What you can do is separate what we observe consistently from what we can reasonably attribute to a cause, and work on the factors that stay under your control: clarity, consistency, coverage and external confirmation.
The biggest risk is not picking the wrong tactic, it is mistaking a coincidence for a rule. That is why a repeatable test protocol has to come before any conclusion.
Taxonomy: observation, correlation, causality
| Level | Example | Correct use |
|---|---|---|
| Observation | Answers often cite institutional sources and consistent sites | Sets priorities |
| Correlation | Businesses with clean entity data appear more often | Hypothesis to test |
| Causality | This change produced the citation | Rarely provable |
| Speculation | There is a trick to being cited | Ignore |
What is consistently observed
- Content that answers directly is picked up more than promotional copy.
- Consistency across sources reduces inaccurate descriptions.
- Technically inaccessible pages do not appear.
- Geographic and category specificity helps with local questions.
Prompt variance: why a single test is worthless
Changing one word in a prompt can change the list of businesses cited. Even the same prompt repeated can return different results. Any conclusion based on an isolated test is, statistically, an anecdote.
Repeatable test protocol
- Define 15–25 fixed prompts representing real buying questions.
- Repeat each prompt at least three times, in clean sessions.
- Log presence, position in the answer, description and cited sources.
- Repeat the cycle on a fixed cadence, for example every 30 days.
- Introduce one substantial change per cycle and observe the effect.
Note
Log the competitors cited too: it tells you more about the market than any theoretical analysis.
What stays under your control
You do not control the model. You control the quality and consistency of the information the model can find, the coverage of buying questions, and your presence on the sources assistants tend to use. That is where budget belongs.
FAQ
- Is there a way to guarantee citation?
- No. Anyone guaranteeing it is not describing how these systems actually work.
- How often should tests be repeated?
- Monthly for stable monitoring, or after any substantial change to your online presence.
- Does structured data help?
- It helps entity understanding and remains useful for traditional search. It is not a magic lever.
- Lots of content or a few strong pieces?
- A few genuinely decisive, mutually consistent pieces almost always beat a high volume of similar pages.
Sources
Related intelligence
- AI VisibilityHow to Get Your Business Found and Recommended by ChatGPTThe five-level model — eligibility, entity clarity, answer coverage, corroboration, measurement — that decides whether AI assistants consider your company when a buyer asks for a supplier like you.
- AI VisibilityAI Visibility, GEO, AEO and ChatGPT SEO: What the Terms Actually MeanAn objective → discipline → activity → measurement → limitation matrix to cut through acronyms and decide what your business actually needs.
- AI VisibilityHow to Get Customers from ChatGPT: From AI Discovery to Qualified PipelineThe prompt → mention → visit → conversion → pipeline chain: where it breaks and how to close each link with a real commercial path.