AI search optimization (GEO)

Some of the searches that used to go to Google now go to a chat. The decisive difference is the format of the answer: traditional search returns ten links and leaves the choice to the person; an AI assistant returns one answer and cites one or two sources.

That changes the nature of the competition. You’re no longer competing for a position in a list. You’re competing to be the source.

What GEO is

GEO, or Generative Engine Optimization, is structuring content and data so that language models find, understand and cite your company. It covers ChatGPT, Perplexity, Gemini, Copilot and the AI-generated answers inside Google Search itself.

It shares its technical foundation with technical SEO: if the crawler can’t read your page, the model can’t either. But the target is different, and that changes how content is written and marked up.

How it differs from traditional SEO

	SEO	GEO
Goal	Position in the list	Being the cited source
Unit that matters	The page	The passage that answers
Format that wins	Comprehensive content	Direct, verifiable statements
Authority signal	Backlinks	Entity consistency across sources
How it’s measured	Rankings and impressions	Citation rate per question

What works in practice

Answer the question in the first sentence

A model extracts the passage that answers directly. Copy that takes three paragraphs to get to the point yields a poor passage. The structure that works is claim first, reasoning second. The opposite of copy that sets the context first.

Declare your entity consistently

The model needs to recognize your company as one specific thing in the world. That requires your name, address, CNPJ (Brazil’s company registration number), field of business and social profiles declared the same way everywhere: on your site, in your structured data (in Portuguese), on Google Business, on LinkedIn.

Inconsistency breaks the association. A different address in the footer and in the schema is enough for the model to treat the two mentions as separate entities, and neither one builds authority.

Give the data, not the adjective

“We’re an industry leader” isn’t citable. “Core Web Vitals with LCP under 2.5 s on every page we deliver” is. Models cite specific claims, with a number, a timeframe or a criterion. Something that can be repeated without turning into advertising.

Say what you don’t do

Counterintuitive, and it’s what works best. Content that spells out where a solution doesn’t fit is treated as more trustworthy, and it gets cited in answers to comparison questions. Those are exactly the questions people ask when they’re making a buying decision.

Serve an llms.txt

A plain-text file at the root of your domain that describes the company, its services and its main content in direct language. It’s an emerging standard (in Portuguese), not yet universally adopted by the models, and cheap to maintain. This site serves its own at /llms.txt (in Portuguese).

How we measure

By asking. There’s no Search Console for AI assistants, so measurement is done by hand, with a method.

1. We build a set of questions a real customer would ask before hiring, from generic to comparative.
2. We run that set on the main assistants, in clean sessions, at regular intervals.
3. We record which questions cite you, who gets cited instead of you and from which source.
4. We compare over time. The metric is the citation rate per question set.

It’s labor-intensive and has variance, since model answers aren’t deterministic. But it’s measurable, and it’s the only honest way to know whether the work is having an effect.

What we don’t promise

There’s no way to guarantee a citation. Models change versions, change data sources and don’t publish their selection criteria. Anyone offering a guarantee that you’ll show up in ChatGPT is selling something they don’t control.

What you can do is raise the odds, with content that answers well, an entity declared consistently and verifiable data. And measure, so you know whether it worked.