What changed
For twenty-five years, search returned a list and the user chose. Increasingly, an AI system reads several sources and writes one answer — in a chat assistant, in an AI overview above the results, or inside a tool someone is already using.
The consequence for a site: you can influence an answer without receiving a visit. Being the source a model draws on, and being named while it does, becomes its own objective. That objective is what GEO refers to.
What tends to get used
A system assembling an answer needs passages it can lift, compress, and attribute without introducing an error. That preference shapes what kind of writing gets drawn on.
| Hard to quote | Easy to quote |
|---|---|
| "Our platform helps teams move faster" | "Itiner drafts a day-by-day itinerary that keeps the agent's own edits when it regenerates" |
| "Pricing is flexible" | "Plans start at ₹999 per month for one user" |
| "Founded some years ago" | "Navital Labs was founded in 2025" |
| A claim spread over four paragraphs | A claim stated once, completely, in one paragraph |
- Self-contained paragraphs — a passage that only makes sense after reading the previous three won't survive extraction
- Direct answers under question-shaped headings
- Concrete nouns, numbers, and dates rather than adjectives
- Named entities — say the product name and the company name rather than "we" and "it"
- Explicit dates, so a system can prefer current information
Why attribution and consistency matter
A system choosing between conflicting sources looks for corroboration. If your pricing says one thing on your site, another on a directory, and a third in an old blog post, none of the three is safe to state, and the confident competitor gets quoted instead.
Named authorship helps for the same reason. A lesson with a real author, a role, and a page describing their work is easier to attribute than an anonymous article — and attribution is the currency here.
What you can and can't measure
Be honest about this. Generative answers are personalised, non-deterministic, and mostly don't send a referrer. There is no rank position to check, and the same question asked twice can return different sources.
- You can ask several assistants the questions your customers ask, and record what they say and who they cite — repeated over time, that's a real signal
- You can watch for a rise in direct traffic and branded search, which is often where this shows up
- You cannot get a reliable rank number, and any tool promising one is estimating
Applying it
- List the ten questions a customer asks before buyingIn their words. These are the prompts, not the keywords.
- Give each one a page or a clearly headed sectionQuestion as the heading, answer in the first paragraph beneath it.
- Replace vague claims with checkable factsNumbers, dates, names. Everything a system could get wrong by paraphrasing.
- Make every important fact consistent everywhere it appearsSite, directories, profiles, older posts. Fix or remove the stale ones.
- Ask the assistants your ten questions, monthlyLog the answers. The trend is the measurement.
Common mistakes
- Treating GEO as separate from SEO — uncrawlable pages can't be cited
- Writing to be quoted rather than to be accurate
- Leaving contradictory facts across your own properties
- Expecting a rank number that doesn't exist
- Burying the answer three paragraphs into a page
Key takeaways
- Generative answers mean influence without a visit, so being the cited source is the goal
- Self-contained, specific, named, and dated passages are the ones that get used
- Consistency across your properties decides whether a fact is safe to repeat
- Measurement is directional — log assistant answers over time rather than chasing a rank
Try it yourself
Ask two AI assistants a question your customers ask before buying. Note whether you appear, who does, and whether anything said about you is wrong. That wrong thing is usually traceable to a stale page you'd forgotten about.