What generative engines actually cite, and why
Strip the branding off AI Overviews, ChatGPT's browsing mode, Perplexity and Copilot and they share one shape: a query gets rewritten into several sub-queries, each hits a search index, a handful of pages come back, and the model writes an answer out of what it can extract from them. The citation is a receipt for that extraction.
That mechanic explains almost everything about which pages get picked. It isn't mystical and there's no new algorithm to reverse-engineer.
Cited pages share four properties: they were already retrievable for the sub-query, they hold a passage that answers it without the rest of the page, they say something specific enough to quote, and they name entities the model can resolve. A page ranking fifth that holds the answering paragraph beats the page ranking first that buries it.
The deliverables that change, and the ones that don't
Most GEO pitches are a re-skin of an SEO scope of work. Here's the line-by-line, so you can see which parts you'd genuinely be paying extra for.
| Deliverable | Under SEO | Under GEO | Really different? |
|---|---|---|---|
| Crawling and indexing | Required | Required — retrieval runs on the same index | No |
| Keyword-mapped pages | The core of the plan | Still the core; queries are longer and more conversational | Barely |
| Answer formatting | One snippet target near the top | A self-contained 40–60 word answer under every H2 | Yes |
| Original statistics | Useful for earning links | The most quotable unit there is — a model lifts a number and credits the source | Yes |
| Comparison tables | Good for featured snippets | Rows get lifted close to verbatim on "X vs Y" questions | Yes |
| Entity consistency | Rarely audited by anyone | Name, category, founders, location and pricing must agree everywhere | Yes |
| Schema markup | Rich-result eligibility | Helps entity resolution. Not a citation switch, and nobody has shown it to be one | Partly |
| Backlinks | Authority and discovery | Same, plus the unlinked brand mentions that build association | Partly |
| Core Web Vitals | Ranking input and real UX | Almost irrelevant to a model reading extracted text | No — keep doing it |
llms.txt | Not applicable | Proposed, adopted by no major engine. Twenty minutes to add; expect nothing | No |
Entity consistency is the one genuinely new discipline
The other three changes are formatting habits a good editor picks up in a week. Entity consistency is different work, and most teams have never done it.
A generative engine has to decide that the business named in a Bengaluru directory, the one quoted in a trade publication and the domain it just crawled are the same organisation. Everywhere your details disagree, that confidence drops — and a model unsure who you are will leave you out and cite someone unambiguous.
This is unglamorous audit work: one spreadsheet, every surface where your business is described, a column per fact that has to match.
- Legal and trading name — one form, everywhere. "Pvt Ltd" in some places and not others is a real ambiguity, not pedantry.
- Category description. If your site says "SEO agency", LinkedIn says "growth partner" and a directory says "web design company", you've described three businesses.
- Founders and named people — same spellings, same roles, on an about page with
sameAspointing at their real profiles. - Location and service area — one address format, and an explicit statement of where you work.
- Pricing and claims. If your site says ₹75,000/mo and an old press piece says ₹25,000, the model can't tell which is current.
- Organization schema on the homepage, with
sameAsfor every profile you control. The cheapest disambiguation signal available.
How to measure this when no rank tracker covers it
There's no rank tracker here, because there's no stable position-based result to track. Answers vary by user, session, phrasing and day. Three honest routes exist, each with a hole in it.
- Search Console, read for the signature. Google folds AI Overview impressions and clicks into the standard performance report rather than splitting them out, so you can't isolate them. What you can watch is the pattern: impressions holding or rising on informational queries while CTR falls. That's a pattern, not proof — say so when you report it.
- GA4 referrals from assistant hostnames. Traffic from
chatgpt.com,perplexity.ai,copilot.microsoft.comandgemini.google.comarrives as referral. Small volumes, usually high intent. Build a custom channel group so it stops hiding inside "Referral" or "Unassigned". - A manual prompt log. Fix a panel of 25–30 prompts a real buyer would type. Run them monthly, same day, logged out, same engines. Record whether you were cited, which page, and who else was in the answer. Tedious, a sample rather than a census, and still the most truthful number here.
The verdict, and what to do on Monday
Forced to pick one, pick SEO — not out of caution, but because GEO without indexing and rankings is a scope of work with no distribution attached. Every generative surface reads from a search index, and the index doesn't care what you call your discipline.
If you already rank well and you're watching CTR erode on informational queries, that's the moment GEO earns its 15%. Reformat before you write more. Then publish one number nobody else has.
Monday's list: take your top 20 pages by impressions, add a 40–60 word answer under every H2, fix Organization schema and sameAs, start the prompt log. A week of work, covering most of what GEO means in practice. Longer argument in is generative engine optimisation real.