Generative Engine Optimization
GEO — Generative Engine Optimization — is the set of choices that make a website understandable, extractable and quotable by a generative engine such as ChatGPT, Claude, Perplexity, Gemini, Google AI Mode or Copilot, when it composes an answer instead of returning a list of links.
In one line: SEO works to get your link into a list the user scrolls. GEO works to get your sentence into the answer the user reads.
They do not replace each other. Most of the groundwork is shared, and someone who does not find what they need inside a generated answer still clicks through. A site has to do both — which is why Timbaly does not sell them as two products.
The difference, concretely
A search crawler is deciding where to rank a document. A generative engine is deciding which sentence to lift and whom to credit for it. Those two jobs reward different writing.
| Classic SEO | GEO | |
|---|---|---|
| Goal | Rank the page in a list | Have a passage quoted and attributed |
| Unit that competes | The page | The paragraph, sometimes the sentence |
| Rewards | Authority, links, intent match, speed | Self-contained answers, clear entities, verifiable specifics |
| Punishes | Thin and duplicated pages | Vagueness, hedging, claims with no number attached |
| Feedback | Impressions, clicks, position — in a console | None published by any engine |
| Time to effect | Weeks to months, then durable | Fast when the model retrieves live; opaque when it answers from training |
The last two rows are the ones that get glossed over in most GEO sales pitches, and they are the two that decide what you can honestly promise a client.
The checklist
This is the list Timbaly applies to every page it writes. None of it is exotic; the difficulty is doing all nine on every page, every time, which is exactly the part a platform is better at than a person.
The first forty to sixty words under the heading answer the question the heading asks, with no back-references. A model extracting one paragraph has to find it complete. “As we saw above” makes a passage unusable the moment it is lifted.
“How much does it cost?” rather than “Pricing”. Those are the strings a model matches and recomposes, and they are also how people actually type.
An engine quotes the sentence, not the table cell. “€19 a month, per site” inside a paragraph travels; the same number alone in a column often does not.
The same business name, the same city, the same service wording throughout. A model that finds three variants of your name has three weaker entities instead of one strong one.
JSON-LD for LocalBusiness, Service, FAQPage, Product, Offer — saying the same thing the page says. Markup that contradicts the copy is worse than no markup.
Not “What are your service modalities”. “Do you come out on Sundays?” The question is half the retrievable content.
A plain-text file that says what the site is and which pages matter. A proposed convention rather than a standard — cheap to publish, easy to remove. What it is and how to write one.
GPTBot, OAI-SearchBot, ClaudeBot, PerplexityBot, Google-Extended and the rest, allowed explicitly rather than by silence. Silence is ambiguous; a named rule is a decision.
If your site says one set of opening hours and your Google Business Profile says another, a careful model cites neither. Consistency across sources is a retrieval signal you control for free.
The part nobody sells honestly
No generative engine publishes a citation metric. ChatGPT, Claude, Perplexity and Gemini expose no console and no measurement API. There is no authoritative number to read, for anyone, at any price.
Reviewed August 2026. If any of these engines ships a real citation report, this page changes the same week.
The only way to produce a figure is to ask the models the same questions repeatedly and record what comes back. That is sampling, not measurement: the answer varies by phrasing, by region, by session, by model version, and it costs money per question asked.
Sampling can still be useful — it tells you roughly whether you show up for the ten questions that matter most to your business. What it cannot do is give you a trend line you should make decisions from, and it should never be sold as one.
Everything that has a real source: Google Search Console (queries, impressions, position, the canonical Google actually chose), Bing Webmaster Tools, PageSpeed on mobile, IndexNow submissions, and first-party analytics for what visitors do once they arrive.
Where a number does not exist, the tools say so rather than returning zero. A model reading zero clicks concludes your site gets no traffic from Google — a false statement said with confidence, which is the worst way an integration like this can fail. See exactly which numbers come from where.
What Timbaly does about it
The nine points above are in the instructions every page is written against, and the deterministic parts — structured data, canonical, hreflang, the root files — are applied by code rather than requested from a model.
The site scans itself against the same criteria and comes back with actions you approve: this page buries its answer, this one has no FAQ, these two compete for the same query.
Queries with impressions but weak positions become the next pages to write, so the plan follows a measurement rather than a hunch.
Questions about GEO
No. They share most of the same groundwork — crawlability, structure, speed, structured data — and they pay off in different places. People who do not find what they need in a generated answer still click a link, and generative engines lean heavily on the same index that classic search built. Treating GEO as a replacement is how sites lose the traffic they already had.
Not from a console, because none exists. Anyone showing you a citation dashboard is asking the models questions and tallying the answers. That can be worth doing, but it is sampling: the result shifts with phrasing, region, session and model version, and it costs per question. It is not a metric with a source you can audit.
It is a proposed convention, not a standard any engine has committed to honouring, and no major provider has confirmed using it as a retrieval or ranking input. It costs nothing, it documents your site in one place a human can also read, and removing it is one deletion. Publish it as cheap insurance, not as a lever — and be sceptical of anyone who bills for it as a service. The full guide, with an example to copy.
It depends on whether you sell attention or content. If you want to be recommended — a plumber, a law firm, a restaurant, a consultancy — then blocking GPTBot or ClaudeBot removes you from exactly the answers you were trying to appear in. If your content is the product, and being summarised for free costs you the visit, blocking can be the correct commercial decision. Timbaly allows them by default and the file is yours to edit.
No, and the reason is mechanical rather than commercial: free sites are served with noindex on a shared platform subdomain. A page that asks not to be indexed will not be retrieved or cited by anything. Free is for seeing what the platform produces; €19 a month is what makes it findable. The plan comparison.
Every page Timbaly writes follows the nine points on this page, and the technical declarations are applied by code rather than remembered by a person.