How presence is built in AI answers.
This is the method we apply to clients, and to ourselves, every week. We publish it because the value is not in the secret: it is in the execution and the measurement. Signed by Eduardo Mendonça, kept alive, with a date on everything that ages. The PDF edition is in Portuguese for now.
Three acronyms, one map
Classic search: Google, Bing. The base GEO is built on.
Appearing in answers generated by AI: ChatGPT, Claude, Gemini, Grok, Perplexity, Copilot.
Optimising for the direct answer: featured snippets, voice, AI Overviews.
| In SEO | In GEO |
|---|---|
| You optimise for 10 blue links | You optimise for 1 answer |
| The user chooses between results | The AI chooses for them |
| Keywords and backlinks | Structure, authority and citability |
| You measure positions on Google | You measure mentions in AI engines |
First, understand the two routes
Every engine answers through one of two routes, or a mix of the two. Almost everything written about GEO treats the two as one, and that is where the unrealistic promises come from.
Memory
MonthsThe model answers with what remained from training. It does not go online and cites no link. The winner is whoever is a recognised entity across many independent sources.
Levers: Entity, durable press, original data cited by third parties.
Nothing you do today changes this route tomorrow. Entity and authority work is sowing for the next training cycle, and whoever does not sow stays out for years.
Live search
WeeksThe question triggers a search, the engine reads pages at that moment and answers citing them. The winner is whoever has findable content that answers the question.
Levers: Comparative content, healthy indexing, presence in the sources the engine prefers.
This is where the first results show, and where the answer changes without warning: the page the engine reads today may not be the one it reads next week. Hence continuous measurement.
How an engine decides whom to cite
Four steps between the question and the answer. Each dimension of the method exists because it moves one of these steps; when an action moves none of them, it is not GEO, it is busywork.
The question is classified
The engine decides whether to answer from memory or to search. Factual, stable questions tend to stay in memory; questions with a price, a date or a comparison trigger a search. The same brand can exist on one route and not on the other.
The search fans out
One question becomes several internal queries. The engine does not search the user's sentence: it searches variations, synonyms and sub-questions. Doing well on one phrasing and badly on its neighbours is normal, and measurable.
You compete at the passage level
The unit the engine extracts is not the page, it is the passage. In Google's AI Overviews the answer is typically in the first sentence of the extracted passage. An excellent page with the answer buried in the fifth paragraph loses to an average page that answers in the first line of the right section.
The synthesis picks the brands
The engine writes with two or three sources inside. This is the step that separates citation (the source with a link) from mention (the name inside the text). Both count, both are measured, and they do not travel together: some are cited without being mentioned, and the reverse.
The eight dimensions of the work
The market’s most common mistake is reducing GEO to technical work on the site. The site is one dimension in eight. A plan that does not touch the other seven leaves most of the result on the table. For each dimension: the thesis, how to check it in minutes, the typical actions, and the mistake we see most often.
Technical foundation
The site has to be machine-readable before it can be convincing to people.
Does robots.txt block GPTBot, ClaudeBot, PerplexityBot? Does the HTML arrive rendered from the server, or does the content only exist after JavaScript runs? Do product and service pages carry the right Schema.org type?
- Allow the AI crawlers in robots.txt: the best effort-to-effect action in the whole method
- Publish llms.txt with real context about the company, not just a list of links
- Schema.org on the right pages: Organization, Person on authors, Product with complete attributes, FAQPage where there are questions
The common mistake: Stopping here. The technical base decides whether an engine can read; it does not decide whether it wants to cite.
Content and topical authority
An engine cites whoever adds something. Interchangeable text gives it no reason to choose you.
If you cover the logo, could the text belong to a competitor? Is there a single page that answers the category's comparative questions head-on, including the ones that name competitors?
- Topic clusters with real depth, not loose posts chasing keywords
- Your own data and statistics: the raw material third parties cite, and engines after them
- Honest comparative content, including where the competitor is better: it is the format engines reuse most
The common mistake: Cadence without substance. Publishing a lot of interchangeable content moves no engine.
Entity and brand
The model does not store pages, it stores entities. A brand only asserted by itself is a claim without confirmation.
Does the brand have a Wikidata item? Does the site's sameAs point to the same profiles everywhere? Does searching the company name on Google show a Knowledge Panel?
- Create and maintain the Wikidata item, after the public references exist: free, within reach of any brand, and the entity source models consume most
- Wikipedia when there is notability for it (and no forcing when there is not)
- Name, address and contact exactly the same in every directory and profile
The common mistake: Treating it as one-off work. A misaligned entity drifts again; it is reviewed every quarter.
Authority and digital PR
The sources engines cite are frequently third parties. Being in them is worth as much as your own site.
When an engine answers about the category, which sites does it cite? Does the brand exist in any of them? How many of those sources mention competitors and not the brand?
- Reference press for the category and the country: not any clipping, the sources engines actually cite
- Podcasts and conferences: transcripts are indexable text with the brand name in a context of authority
- Anchors with the brand name instead of “learn more”: the link text teaches the engine to associate name and topic
The common mistake: Buying link volume like it is 2015. The engines' citation graph rewards the right sources, not many sources.
Social and community signals
Engines cite community platforms with a weight that surprises anyone coming from classic SEO.
Does the brand exist on Reddit and in the forums where the category is discussed? Is there YouTube video answering the category's questions? What do the threads say where the brand appears, and the ones where it does not?
- Genuine presence in the category's communities: real participation, not brand spam, which communities expel and engines learn
- YouTube with answers to buyer questions: it leads citations in some engines
- LinkedIn for B2B and GitHub when there is a technical product
The common mistake: Treating community as a publishing channel. It is third-party conversation; you earn it, you do not post it.
Authority on your own site
Whoever writes has to be someone, and that someone has to be verifiable. The marked-up author and the visible author must match.
Do articles have a named author, visible on the page, with Person schema and sameAs pointing to real profiles? Do case studies have verifiable results or adjectives?
- Declared authors with Person + sameAs, and the same name visible in the text; schema without the visible name is half the job
- Named credentials: who the person is, why they know the subject
- Cases with numbers and dates instead of superlatives
The common mistake: Anonymous institutional content. A method without an author is exactly what an engine has no reason to prefer.
Measurement and feedback
Without repeated measurement you cannot tell whether the work paid off, and you argue from impressions.
Can anyone in the company say, with a date, in which engines the brand appears today and in which it does not? Is there a history, or only the memory of a search someone ran one day?
- Real buyer questions run week after week, in the same engines, under the same conditions
- Search Console and GA4 with the AI channel separated: a click from an answer is not a click from a result
- Bing Webmaster Tools, because Bing's index feeds ChatGPT and Copilot
The common mistake: Measuring once and concluding. The answer changes without warning; the measurement only means something as a series.
Strategic positioning
Appearing is half the way. The other half is appearing for the right reasons, in the questions that make money.
Of the questions where the brand appears, how many are bottom of funnel? Who appears in the category's purchase-intent questions? Are there valuable questions where nobody appears?
- Share of voice against peers, question by question, not on average
- Map the questions to the funnel and concentrate the work where the answer influences money
- Occupy open territory: questions with no owner are the cheapest acquisition there is in GEO
The common mistake: Optimising for the highest-volume question instead of the highest-intent one.
Each engine rewards different work
The practical guides, with verified commands, are in Portuguese for now: ChatGPT → · Claude → · Gemini → · Perplexity → · Copilot → · Google AI Mode →
Commerce: products inside the answers
For anyone selling products, the AI answer is not just text: it is text, carousel and buying surface, and the three diverge. The base work is the same (readable product pages, one product per page with its schema, a complete Merchant Center feed), but the measurement has to look at each surface by name.
The answer in text
The engine recommends product families and brands in prose. It lives on the two routes like everything else: memory for reputation, search for the current catalogue.
The shopping carousel
The ChatGPT app shows product cards with price and store, built from the pages the search retrieved at that moment. It is not a paid feed: it is the catalogue of whoever has readable product pages. And it diverges from the text: in our August 2026 measurements, with 13 products across three surfaces, there was not a single exact overlap.
Google's and Microsoft's surfaces
AI Overviews, AI Mode and Copilot each have their own product behaviour, and none replicates the others. Whoever measures only one is seeing a third of the picture.
The service: Agentic commerce → · the full chapter, product by product, in Portuguese: Products in AI →
Ads in AI engines
Earned presence is the centre of the method, but paid enters the same board, and whoever manages the two together decides better. The state per engine, verified as of the date:
The management service, with budget and scope: AI ads →
What is measured, and how
Measurement is the dimension that validates the other seven. We run real buyer questions, every week, across eleven engines and surfaces, always under the same conditions, because only the series lets you say the work paid off.
Citation rate
In what fraction of questions the brand appears, per engine. Not a single average: you see in which engine you win and in which you disappear.
Share of voice
Who appears in the brand's place, by name, question by question. Real and adjacent competitors kept apart: a classic SEO consultant is not a peer of a SaaS.
Cited sources
Which sites the engines use to talk about the category. It is the map of the authority work: the sources worth being in have names.
Two conditions per question
Each question runs with and without live web search, because the answers diverge between routes, and the difference between the two is information, not noise.
This is what the Visibility Tracker → delivers continuously, with history and alerts.
The plan has four horizons
Because the routes have different rhythms, the plan cannot be a single list. Each action lives in the horizon where it produces its effect, and each horizon mixes dimensions; it is never just one.
Quick wins
What you fix without waiting for anyone: crawlers, schema, product data, the pages that answer questions already being asked.
Technical + content
Foundation
Entity registered, comparative content published, first conversations with the sources the category cites.
Entity + content + authority
Authority
Press, podcasts, your own data published. It is the work the memory engines will absorb in the next cycle.
Authority + social + positioning
Operation
Weekly measurement, response to what changes, defence of the territory won. GEO is not a project with an end; it is a position you hold.
Measurement + everything else
Five wrong ideas that cost money
“GEO is a trick you apply.”
There is no trick: there are eight dimensions of work and measurement to know what paid off. Whoever sells a shortcut is selling what they do not command.
“Immediate results in every engine.”
The two routes have two rhythms: weeks on live search, months in memory. Promising the same timeline for both is not understanding the mechanism.
“A technically perfect site is enough.”
The site is one dimension in eight. Third-party sources, the entity and the community weigh on what the engine answers, and they sit outside the site.
“One measurement is enough to know where we stand.”
The answer changes without warning, between engines and inside the same engine. A snapshot says little; the series is what says whether the work is paying off.
“What the API returns is what the user sees.”
The app and the API diverge: the shopping carousel does not even exist in the API. You measure the surface the buyer uses, not the one that is convenient to measure.
The rules we do not negotiate
Numbers instead of adjectives. “Citation went from 4.6% to 7.1%”, never “improved a lot”.
Statistics only with a source and a date. If there is no study, we say there is no study.
Vendor data treated as direction, not as fact.
First-hand verification before publishing about launches: we go as far as the thing stops working.
Every recommendation ends in an action with an effort estimate. Diagnosis without a plan is entertainment.
When the work does not pay off for a client, we say so. There are criteria for when GEO is not a priority.
Frequently asked questions
+Why publish the method?
Because the value is not in the secret; it is in the execution and the measurement. Anyone can read these eight dimensions; few organisations have the discipline to work all of them, every week, with the numbers in front of them. Publishing it also keeps us honest: what is written here is what we do.
+Does GEO replace SEO?
No. GEO sits on top of SEO: AI engines read what is indexed and served, and a site with a weak technical base performs poorly in any engine. The difference is the target. In SEO you optimise for ten links the user chooses from; in GEO you optimise for one answer in which the engine chooses for the user.
+How long until GEO shows results?
It depends on the route. In the engines and modes that search the live web, weeks: as soon as the content exists and can be found, it can be cited. In those that answer from memory, months, because they depend on training cycles the vendor decides. Anyone promising the same timeline for both has not understood the mechanism.
+Is this playbook enough to do the work alone?
To start, yes: the checks in each dimension are described and any team can run them. What the playbook does not replace is the weekly measurement across eleven engines and surfaces, under identical conditions, which is what tells you whether the work paid off. Without the series you are arguing from impressions.
The method applied to your brand starts with a measurement.
We run hundreds of real buyer questions across eleven AI engines and surfaces and show where you appear, who appears in your place, and which of the eight dimensions to start with. Free and without commitment.