How AI search decides which sites to name.
An AI answer names a few sources, and a page can pass three of four checks and still never appear. This guide shows the four stages, how to tell which one you are failing, and a real example from Mindstorm Labs' own results. Mindstorm Labs applies the same method in its AEO and GEO consulting.
What decides whether an AI answer names your site?
Four things, in order: whether the system can fetch and read your page (eligibility), whether your page is a candidate for the question or its sub-questions (retrieval), whether a passage on it answers the question on its own (extraction), and whether independent sources confirm who you are enough to name you (synthesis). The first question you answer "no" to is the stage to fix.
This four-stage model is Mindstorm's working framework from the forthcoming book The Art of Being Found Online. It is a diagnostic scaffold, not a description of any vendor's internal system.
What does each stage check, and how does it fail?
| Stage | The system asks | What you control | Typical failure |
|---|---|---|---|
| A. Eligibility | Can I fetch, render and use this page? | robots rules, crawler access, rendering, status codes, indexing | Blocked or unrendered, so the page never enters the pool |
| B. Retrieval | Is this page a candidate for the query or its sub-queries? | Topical coverage, sub-query coverage, organic strength | Strong on one head term, absent on the sub-questions |
| C. Extraction | Which passage answers this, and is it self-contained? | Passage structure, answer placement, entity clarity | Retrieved, then skipped because the answer is buried or vague |
| D. Synthesis | Which sources do I trust enough to name? | Third-party corroboration, entity consistency, evidence | Used silently, or replaced by a source with more outside confirmation |
You control the first three on your own site. The last also depends on sources you do not own, which is why off-site mentions matter.
How do I tell which stage I am failing?
Build a fixed prompt panel of 30 to 50 prompts that mix informational, comparative, local and transactional questions. Run each on the engines your buyers use at least three times, since answers vary. Log whether you were cited with a link, mentioned by name, or absent, and whether what was said about you is correct. Then read the result like this:
| What you see | Likely stage | First action |
|---|---|---|
| Page is not in the source list and fetching is blocked | A | Fix access before anything else |
| Ranks on page one but is never cited | C or D | Rework passages, then audit corroboration |
| Does not rank for the head term or its sub-queries | B | Coverage and authority work |
| Named, but with wrong or stale facts | D | Reconcile entity data across properties |
| Cited on one engine, absent on others | A or B for the absent engines | Check each engine's crawler access separately |
What did this look like for Mindstorm Labs?
On October 13, 2026 we ran 19 prompts, three times each (57 runs), on Gemini with Google Search grounding. We counted whether mindstormlabs.dev was cited and whether the Mindstorm Labs or Jeffrey Jenson name was mentioned.
| Prompt group | Runs | Our domain cited | Our name mentioned |
|---|---|---|---|
| Branded ("Who is Jeffrey Jenson of Mindstorm Labs?") | 12 | 11 of 12 | 12 of 12 |
| Local ("AI consultant Cedar City Utah" and similar) | 21 | 0 | 0 |
| National ("AEO GEO consultant for small business" and similar) | 24 | 0 | 0 |
Diagnosis. Stage A passes: the site is crawlable, indexed and cited whenever someone asks about the brand by name. Stage B fails for every unbranded question, because no page of ours is a candidate for those queries yet. Only one of 19 prompts gave a different result across its three identical runs, so the pattern is stable on this engine. The first fix is coverage: pages that directly answer the unbranded local and national questions.
Do GEO and AEO tactics work?
- The Princeton-led GEO paper (Aggarwal et al., arXiv 2311.09735, KDD 2024) reports visibility gains of up to 40% on its benchmark, with the largest effects from adding citations, quotations and statistics. Its main table uses a single GPT-3.5-based engine, and a benchmark result is not a production guarantee.
- Google's guide to generative AI features says optimizing for generative AI search is still SEO, that no special markup or new files are required, and that it treats AEO and GEO as industry terms. We read that page as updated July 10, 2026.
- For engines other than Google, passage-level tactics are hypotheses to test in your own panel, not established rules.
Evidence labels used across this work: Primary (read from the source), Reported by others (original not retrieved), Practitioner (a named operator's claim), Model (Mindstorm's working framework).
Where does this come from?
Our AI Visibility Index study looked at which sites Gemini cites for 300 local-service queries and found that schema and FAQ blocks did not predict being cited, while Reddit and review sites dominated. The method, data and code are public on GitHub. The book chapters behind this guide are in progress: see the books.
If you want this done for your business, see AEO and GEO consulting for small businesses.