SEO, GEO and AEO optimization

The system splits the question into several queries, collects dozens of pages and builds the answer from short fragments. It is the fragment that gets quoted, not the page as a whole. The first requirement: every important paragraph has to be understandable on its own - a paragraph that starts with “as I mentioned above” says nothing when cut out.

The second requirement is numbers. “An individual approach to every client” is not quotable because it contains no information; “a full SEO audit of up to 5 pages costs 300 EUR and takes 3-5 business days” is quotable because it can be retold without changes.

The third requirement is entity clarity: who you are, what you do and where you work, and this information has to match on the website, in Google Business Profile, on LinkedIn and in directories. If on the website you are a designer and developer but in a directory a “marketing agency”, the model chooses the clearest source, and it is usually not yours.

robots.txt determines access. I often find that a site is closed to AI bots by accident: a security plugin, a Cloudflare rule or a hosting switch called “AI protection” did it. The first job in every audit is to check whether the bots are allowed in at all.

llms.txt is a proposal, not a standard: a file in which a site lists its most important pages in a machine-readable form. No major model maker has confirmed reading it, so I prepare it - half an hour of work that breaks nothing - but I do not promise a result from it.

Boring things have a bigger impact than llms.txt: a valid sitemap.xml, correct HTTP status codes, canonicals without contradictions and pages that open in under two seconds on a mobile network.

The format is described by the llms.txt specification, and each bot's behavior by its owner's documentation: OpenAI, Anthropic and Perplexity.

Schema markup is a JSON-LD block in the page code that tells the machine what a person reads from the design: this is a service, this is its price, this is a question and this is the answer. The types are maintained by schema.org; the ones Google actually uses are listed in the Google structured data documentation.

Schema is not a ranking factor, and I do not promise it as a rankings booster. It reduces guessing: the less the parser has to guess, the greater the chance that the fragment is used in the right context and with the right price. The same applies to entity clarity outside the code, which is what brand identity sorts out.

In 2023 Google switched off FAQ rich results for most sites, so the expandable questions no longer appear in the search engine. I keep the structure anyway, because AI parsers read it.

Queries like “best dentist in Riga” or “door installation near me” bring clients who are ready to buy today, and the work list for these queries is different.

One warning: I do not make dozens of nearly identical pages with a swapped city name. Google calls these doorway pages and lowers their rankings, and AI models ignore them. A local page is valuable only with real content about the specific place: completed sites, delivery times, prices and a contact person.

Models answer local questions based also on Google Business Profile, directories, forums and reviews. If there is nothing about you there, a polished website alone will not change the answer.

The same list repeats in audits regardless of the industry and of how expensive the website was. Most of it can be fixed in two weeks; it gets expensive when the content has to be rewritten and the rendering rebuilt.

ChatGPT has no Search Console of its own, and no tool shows the full picture. Promises of “full AI visibility control” are sales tricks.

The most important tool is a set of questions: I record whether the brand appears in the answer, in which position and which domain is quoted. Part of this tracking is taken over by AI agents for content and measurement. Google Search Console does not show impressions from AI Overviews separately, so an exact number cannot be obtained from its data.

In manufacturing and construction: Tenter, ARM Metals, Estire and Improvement. In logistics and business services: MLM Cargo, Box Latvia, Universal Solutions and DarbaGuru. In retail and consumer services: Tavasdurvis, Lucky Punch, Cafeteria, Valis Stomatoloģija, Lauvas Zobs and Forevolt. In digital products: Apmeklē.lv, Digitālais Dzintars and LimeBit. The published work is in the website and logo design examples.

In eighteen years I have worked with clients in Latvia, Estonia, Lithuania and the USA. Some projects are new websites with the SEO foundation laid during development, some are cleanups of existing sites. The industries differ, but the first problems are the same: content bots cannot read, prices named nowhere, and schema that is either missing or wrong.