Citations, not positions
There is no position five in a generated answer. A source is either drawn on or it is not, and the set changes between assistants and between runs of the same prompt. Treat any single answer as one sample, never as a rank.
This breaks the habits that classic rank tracking built. A position is a stable number you can watch move; a citation is a binary outcome from a process with real variance in it. The same prompt, asked twice an hour apart, can produce different sources without anything about your page having changed.
| Classic search | Generated answer |
|---|---|
| Ordered list of results | A set of sources, unordered |
| Position is stable enough to track | Inclusion varies between runs |
| One ranking per query | Different answers per assistant |
| Click is the outcome | Being quoted is the outcome |
What makes a page quotable
Retrieval-based assistants weigh a page's opening heavily, so the first passage should answer the question outright rather than lead up to it. An introduction that spends three sentences establishing context before reaching the point gives the model three sentences of nothing to quote.
Specific, attributable claims survive summarisation; vague positioning language does not. "Sitemaps allow 50,000 URLs per file" can be lifted into an answer intact. "We offer industry-leading sitemap solutions" cannot be lifted into anything, because it asserts nothing a reader could check.
- State the answer, then explain it. The explanation is for the reader who stayed; the answer is for everyone.
- Name numbers, limits and dates where they are real. They are the parts a model can carry across without distorting them.
- Keep each claim self-contained enough to make sense outside the paragraph it lives in.
- Attribute what came from somewhere else, so your page is a usable source rather than an unsourced echo.
Different products can use different crawlers
Different products can use different crawlers. For Google's AI Overviews and AI Mode in Search, Googlebot is the crawler that matters: Google says robots.txt directives for Googlebot control Search access. Google-Extended applies to some other Google AI uses, not whether a page can appear in Google's AI Search features.
Other assistants may publish separate user agents and controls, so audit each platform against its own current documentation rather than assuming one generic “AI crawler” rule applies everywhere.
- Read your robots.txt as it stands, not as you remember writing it.
- For Google Search, confirm Googlebot can fetch the pages you want eligible for AI features.
- For third-party assistants, check their documented user agents and controls separately.
- Decide deliberately which non-Search crawlers you want fetching your content — this is a business decision, not a technical default.
- Re-check after any infrastructure change; CDN rules and bot filters block agents that robots.txt permits.
Measuring something this noisy
The variance is the central measurement problem, and the answer is the same one that works anywhere else with a noisy signal: fix everything you can, repeat, and compare like with like.
- Write a fixed set of prompts that represent what you actually want to be cited for.
- Run them on a fixed schedule, in the same assistants, without rewording between runs.
- Record which sources were cited, not just whether you were.
- Compare across runs rather than reading any single one.
- Change one thing at a time on the page, then wait a full cycle before judging it.
Recording competitors' citations matters more here than in classic tracking, because the set is the finding. Knowing that three specific sites are cited for your topic and you are not tells you what the model considers a good source for it — which is far more actionable than your own absence.
Where GEO and SEO stop being different
For Google's AI features, the gap is narrower than the vocabulary suggests: Google says the same fundamental SEO practices remain relevant, with no additional technical requirements or special schema needed for AI Overviews or AI Mode.
Other assistants can use different crawlers, indexes and retrieval systems, so GEO should not assume one implementation model applies everywhere. The practical overlap is still strong: make useful content fetchable, easy to discover, specific enough to support an answer, and measure citations as a separate outcome.