Self-contained answers, with the source for every number

GEO is the practice of making content citable inside AI-generated answers, rather than merely rankable in a list of links. SEO competes for a position in a ranking; GEO competes to be the source an answer cites. These are separate wins: a page can rank first on Google and never be mentioned by a single chat.

There is, and the distinction is operational rather than academic. SEO competes for position in a list of links. GEO works on the citability of a text block — statistics with sources, attributed statements, freshness, presence in the ecosystems models consult. AEO structures the page into self-contained questions and answers, so that an entire block survives out of context and can be extracted on its own. In practice the three overlap: the technical SEO foundation remains a prerequisite for the other two.

Because the shift is already measured in behaviour, not projected in a scenario. Organic clicks fell 42% since the AI Overviews expansion, and on queries where an AI Overview takes the screen, CTR drops 58%. That is not seasonality: it is search migrating to synthesized answers. Anyone depending on organic traffic to generate demand is already paying that bill.

The evidence suggests they do, and the mechanism is the shortlist. Buyers no longer browse ten links: they ask, receive an answer with few sources, and treat it as a short list for consideration. If you are not cited, you do not enter that list — and there is no intermediate position where you stay visible.

Three surfaces concentrate buyer questions: the Gemini app and Google's AI Overviews, ChatGPT with search, and Perplexity. Measuring all three matters because the same page can be cited by one and ignored by another: each pipeline has its own criteria for retrieval, reranking and citation. A single-engine diagnosis hides precisely the asymmetry that matters.

It is not dead and it is not a choice between the two. The technical SEO foundation — indexing, crawlability, useful content, links — remains what answer engines require in order to read you at all. What changed is that it stopped being sufficient. The share of AI Overview citations coming from organic rankings rose from 32% to 54% in sixteen months: the SEO foundation is being absorbed, but the citability layer is new and separate.

It depends on the factor, which is why you should distrust a single deadline. Content freshness typically operates on windows of around 30 days. Ecosystem effects — community presence, video, entity authority — typically appear between 60 and 90 days, because they depend on accumulation rather than publishing. What can be honestly promised is knowing where you stand today and measuring the delta each cycle.

The category stopped being a promise and started having revenue: AI visibility monitoring platforms reached millions of dollars in ARR in under a year. The structural movement behind that is the shift from clicks to synthesized answers, which does not depend on which vendor wins.

It does, with a specific advantage. Engines answer in Portuguese and cite Portuguese sources when those exist and are citable. What is still scarce is GEO content written in Portuguese with the rigour the topic demands — which means less competition for citation on Portuguese queries than on their English equivalents.

There is, and the supporting data is the falling overlap between citation and top-10: the share of AI Overview cited pages that ranked in the organic top-10 fell from 76% to 17%. In other words, most cited sources are not the biggest authority in the niche. The strategy for a small player is a flanking one: start with long-tail questions where no competitor has a structured answer, accumulate citations and ecosystem presence, and move up to head questions as the entity gains authority.

There is no single public ranking like Google's, but there is a common architecture: source retrieval, reranking and generation with citation. Each engine implements those stages with its own criteria — which produces the practical effect of the same page being cited by one and ignored by another. Treating this as a single opaque box is the mistake that leads to optimizing for the wrong engine.

The field's reference study is a controlled experiment rather than a correlational one: it measured the effect of specific techniques on citation rate. Those that held up were citing sources, including quoted attributed statements (the paper measured a gain of roughly 22% in adjusted visibility) and presenting concrete statistics with context. The one that made results worse was artificial keyword repetition. That hierarchy of evidence is what separates what is worth investing in from folklore.

It matters as a floor, not a ceiling. High authority helps you get into retrieval, but it does not guarantee citation — and the evidence that most cited pages are not in the organic top-10 shows it is not the deciding criterion. What authority does is reduce the effort required to be considered.

It does, and the effect is correlational at scale: discussion communities rank among the most cited sources in AI answers. The reason is structural — real conversation contains the language buyers use and the kind of comparative detail marketing content rarely has. But participation has to be honest: disguised promotion is punished by moderators, and the signal reaching the model is that of the entire conversation, including the author's reputation.

It is among the most correlated factors in large-scale studies: video presence appeared as the factor most associated with visibility in an analysis of 75,000 brands. The likely explanation is the combination of indexable transcript with a real engagement signal. To extract value, the video needs a clean transcript — what gets retrieved is the text.

They serve entity resolution: they help the model know who you are before deciding whether to cite you. Wikipedia carries disproportionate weight in training data and the Knowledge Graph anchors identity in Google's ecosystem. For most companies the realistic path is not getting into Wikipedia — it is ensuring factual consistency and structured entity data (Organization schema with sameAs) pointing to every official profile.

It does, and recency bias is measurable in LLM-based retrieval systems. The practical effect is that outdated content loses citations even when it is correct. That turns page refreshing from a one-off task into routine: your highest commercial-intent pages need a review cycle with a visible date.

A partial advantage. The share of citations coming from organic rankings rose from 32% to 54% in sixteen months — the SEO foundation is being absorbed. At the same time, overlap with the top-10 fell from 76% to 17%: most cited pages are not in the top-10. Those with strong SEO need the citability layer to convert ranking into citation; those without it can enter by competing directly on specific questions.

Official guidance is consistent and conservative: there is no separate trick for AI features, and the practices that help in Search still apply. Google also documents how to control crawler access, which matters because accidental blocking is a common cause of invisibility. It is worth reading the primary source rather than intermediaries: it is the only place where the information is not interpreted.

ChatGPT search combines its own crawler with a Bing-class index. That means allowing the search crawler is a prerequisite for appearing — and that blocking training and blocking search are different decisions with different consequences. Brands that do not want to contribute to training can allow search and restrict training, aware of the trade-off.

Perplexity runs a staged pipeline with binary gates, including freshness and domain quality. In practice that makes it more sensitive to recent content, and it tends to cite based on fit with the question rather than brand size — which opens room for those with structured answers to specific questions.

There is no evidence it works, and we do not sell it. An analysis of 300,000 domains found no correlation between the file's presence and AI citations, and Google's documentation indicates the format has no effect on its systems. It is a community-proposed standard, not one adopted by the platforms. It costs little to keep, but it should not take space in your priority budget.

The evidence is mixed, and we treat schema as hygiene and entity clarity rather than a primary lever. Studies attempting to isolate schema's effect on citations show inconsistent results, but official direction is that structured data helps systems understand and display content. It costs little, does no harm and resolves the question of who you are — what it does not do is work magic on citation.

It is a structural decline, and three independent curves point to the same cause. Organic clicks fell 42% since the AI Overviews expansion; on queries where an AI Overview takes the screen, CTR drops 58%; and 73% of B2B sites reported significant traffic loss between 2024 and 2025. The timing tracks the expansion of AI features, not the commercial calendar — which rules out seasonality as the main explanation.

The most consistent reports come from publishers and B2B businesses, and the recurring question is not how do I recover my ranking — it is how do I recover the visit that no longer happens. More than 50 publishers have publicly described a scenario of managed decline, with news traffic down around 26%. The useful reading is that the loss is not of position: it is of clicks, and no ranking optimization brings back a click that the synthesized answer eliminated.

It has arrived, on every front. ChatGPT leads digital discovery in the country, ad formats for the category are expanding into the Brazilian market, and GEO content in Portuguese is still scarce. That combination — present demand with a small content supply — is the authority window opening for whoever publishes first.

There is one, and it is what separates this category from a wave of enthusiasm. AI visibility monitoring platforms accumulated millions of dollars in recurring revenue within less than a year of operation. Recurring revenue at that level means budget is being allocated, not merely newsletter interest.

Week 1: unblock and verify AI crawlers in robots.txt, ensure sitemap and indexing, and establish a baseline of where you stand today. Week 2: apply the citability layer to your five highest commercial-intent pages — answer in the first paragraph, statistics with sources, attributed statements, question-and-answer structure. Week 3: refresh outdated content and make the review date visible. Week 4: publish structured entity data and begin presence in one or two ecosystems. Measurement at 60 to 90 days closes the loop.

Write in self-contained blocks. Open the page and each section with the direct answer to the question it addresses, in a sentence that makes sense on its own out of context. Use headings that reproduce real questions, keep one idea per paragraph with explicit names, numbers and dates, and do not depend on images for essential information — text extraction ignores most of them. Engines process content in blocks and cite blocks, not whole pages.

Three work, and all are legitimate editorial content. Citing sources: referencing studies, institutions and primary data signals verifiability. Attributed quoted statements: the reference study measured a gain of roughly 22% in adjusted visibility. Concrete statistics with context: numbers beat adjectives throughout the literature. The red line is honesty — invented statistics and fabricated citations eventually get discovered, and the reputational cost is permanent.

With real contribution, not infiltration. In communities, join discussions where you are the expert, answer in depth and disclose your affiliation when relevant — moderators reward usefulness and punish disguised promotion, and the signal reaching models is that of the entire conversation, including the author's reputation. On YouTube, prioritize honest comparison and demonstration with a clean transcript. On LinkedIn, what works is the same technical content you would publish on your own site.

If you want to be found, allow them — blocking is self-inflicted invisibility. The crawlers that matter are Google's, OpenAI's training crawler, the ChatGPT search crawler (blocking this one removes your site from answers with search) and Perplexity's. Brands with a deliberate no-training strategy can allow search and restrict training, aware of the trade-off. Accidental blocks are common in configurations inherited from privacy plugins, which is why auditing robots.txt is the first item on any checklist.

It can, and severity varies by engine: AI crawlers have limited or costly rendering capacity, and content that only exists after heavy JavaScript execution may never enter the index. The standard mitigation is delivering essential content in the initial HTML, keeping navigation in HTML links and maintaining an updated XML sitemap. The test is objective: compare the text the browser shows with the text a non-rendering client sees.

An entity is the record models keep about your company, and misalignment there is a silent cause of invisibility. The minimum package: complete organization structured data on the site, with legal name, area served, logo and references to every official profile; absolute consistency of name, address and data across all surfaces, because the model triangulates and divergence breeds distrust; and factual coherence between site, profiles and press.

No — and the measured effect is negative. The reference study showed that artificial term repetition worsens results in a generative context: models are trained to recognize and discount manipulative text, and reader experience degrades along with it. The same applies to citation farms, structured data describing non-existent content, and invented statistics. The real shortcut is prioritizing the right pages and the right questions.

Demand four things, and distrust any report that delivers none of them. First: repetition — model answers are probabilistic, so a question run a single time measures nothing; the report must state how many times each question ran and how results were consolidated. Second: traceability — the engine version used and the collection date must be recorded, or January's number is not comparable to June's. Third: structure — output must be validated against a rigid format before entering the database, so that appeared always means the same thing. Fourth: breakdown — an aggregate number tells you nothing to act on; the report must open by engine, by journey stage and by competitor.

Because that is where most buyers ask. The free, fast chat experiences concentrate usage volume, so measuring there means measuring predominant behaviour — and not a caricature of it, since larger models are invoked within those same experiences when the task calls for it. Measuring only the top tier would produce a portrait of what happens for a minority of advanced users, who tend to be precisely the minority that least drives purchasing.

With protocol, not luck. The three defences are repetition with majority consolidation (a mention only counts when it is stable), validating output against a rigid format before it enters the database, and versioning questions and prompts with a hash record — so that comparison across cycles is legitimate. Add to that pinning the engine version and a recalibration procedure whenever the version changes: without it, a model update shows up in the report as if it were market movement.

It is the central metric of an AI visibility report: the share of relevant niche answers in which the brand appears, weighted by position, sentiment and citation. The intuition is that of share of search applied to the generative era. The isolated value says little — what guides decisions is the breakdown: by engine, because it exposes asymmetry between them; by journey stage, because appearing at awareness and not at decision is a different problem; and by competitor, because it shows who the answer is shared with and in what order. And it is the per-factor delta between cycles that becomes prescription, not the raw number.

The calendar should follow the rhythm of the indexes, not commercial convenience. A complete baseline at onboarding; full re-measurement on a monthly cadence; and weekly tracking of a critical question set on paid plans. The content refresh cycle works well on windows of around 30 days, and ecosystem effects appear between 60 and 90 days. Measuring more often than the index changes produces oscillation without meaning.

No — and that honesty is deliberate. No one can guarantee citation in third-party probabilistic systems; anyone promising that is selling what they do not control. What can be guaranteed with integrity is a transparent and public method, stable and comparable measurement, prescriptions based on the best available evidence classified by strength (Tiers A to D) and strategy review each cycle. If your case depends on a position guarantee, no serious vendor will serve — and it is better to know that before signing.

Five questions, and a report answering fewer than five is incomplete. First: does the brand appear in relevant niche answers? Second: in what position is it mentioned — at the start of the answer or at the end of the list? Third: with what sentiment is it presented? Fourth: is it cited as a source, or merely remembered by name? These are different things, and the second is worth more than the first because it signals the engine used your content as its origin. Fifth: which competitors appear in the same answer, and in what order — because visibility is relative participation, not absolute presence.

Because asymmetry between them is information, not noise. It is common for a brand to do well on one engine and be absent from another — a consequence of different retrieval pipelines with distinct freshness and domain-quality criteria. If you measure only one, you optimize for the wrong engine and conclude you are fine when you are visible to only a fraction of your buyers. The per-engine breakdown is what turns a diagnosis into an action priority.

It covers the foundation, not the new layer. The agency optimizes for Google ranking; generative engines use criteria traditional SEO does not work on: text-block citability, ecosystem presence, freshness and entity resolution. Only 17% of AI Overview cited pages ranked in the organic top-10 — ranking well does not automatically become citation. The proposal is not to replace the agency: it is to give it the measurement instrument and the playbook that SEO cannot see.

You can — the right question is at what cost and with what consistency. The method requires running hundreds of questions across multiple engines, with repetition to eliminate the statistical noise of model answers, comparing against competitors and turning gaps into prioritized prescriptions. Manually that means weekly specialist hours and the method drift that makes each report measure something different from the last — which destroys exactly the comparison that justifies measuring at all.

On specific questions, yes — and that is where the opportunity is. With most citations falling outside the organic top-10, the source menu is open, and engines like Perplexity cite based on fit with the question rather than brand size. The strategy for a new site is a flanking one: start with long-tail questions where no competitor has a structured answer, accumulate citations and ecosystem presence, and move up to head questions as the entity gains authority.

They will change, and the method must be designed for that. Two things survive an engine change: the technical foundation that lets you be read (indexing, renderable HTML, structured data) and the citability factors that evidence supports in an experimental context — citing sources, declaring attribution, presenting statistics with context. What changes is the surface, not the principle. That is why measurement is versioned: when the ruler moves, you see the ruler move, not an unexplained earthquake.

Yes, and the reason is architectural. The function a visibility engine answers is universal: given a list of questions, who appears in the answers, in what position and with what sentiment. That does not depend on the niche. What changes from client to client is the vocabulary of the questions — and it is derived from the business, not the engine. An e-commerce enters with catalogue and categories; a B2B SaaS with use cases and integrations; a local clinic with services and city.

The honest framing is demand protection and capture, not a magic multiplier. What can be measured and promised: citable visibility against your own baseline and against competitors, coverage of the highest purchase-intent questions in the niche, and per-engine trend across cycles. What connects that to revenue is that AI answers already close the consideration loop — those not cited do not enter the short list — while the organic traffic that fed the funnel shrinks.

The hype passes; the structural change is measured. The shift from clicks to synthesized answers is a behavioural fact: Google's referrals to publishers fell sharply year over year, organic CTR drops under AI Overviews, and zero-click dominates informational searches. Each new answer surface — assistants, agents, generative search — needs the same sources. The fad would be believing consumers will go back to reading ten blue links.

Because time is asymmetric in this market. The factors that produce citation — ecosystem presence, entity authority, a history of citable content — accumulate, and each month of delay is compound deficit against those who started earlier. On the cost side, implementing now tends to cost a fraction of what it will cost once the market matures. On the operating side, the twelve-month ruler you build is the asset that cannot be bought later.