The AEO questions base

AEO (Answer Engine Optimization) means optimizing your brand's presence inside AI-generated answers (ChatGPT, Gemini, Perplexity, AI Overviews), not in Google's list of links. The practical difference: SEO chases ranking position and clicks, while AEO chases being mentioned, cited, or recommended within the answer itself — often without any click at all.

Practically, yes: all of them describe optimization for AI answer engines. GEO (Generative Engine Optimization) is the preferred term in academic and enterprise circles, AEO in agency marketing; there's no real methodological difference, only a labeling one.

Gartner projects roughly a 25% drop in classic search volume by 2026 as users migrate to chatbots, and Google is showing AI Overviews on an increasing share of searches. On top of that, traffic coming from AI converts on average 4–5 times better than classic organic traffic, because the user arrives pre-qualified.

No. Answer engines feed on the very same ecosystem that SEO builds: indexable content, authority, citations. AEO is a layer on top of SEO, not a replacement for it — the budget gets rebalanced (typically 10–30% toward AEO), not moved wholesale.

A mention means your brand name appears in the body of the answer (a recommendation). A citation means your site is listed as a source or link behind the answer. You can be cited without being recommended, and vice versa; a complete strategy tracks both.

It works wherever customers ask questions before they buy: B2B services, e-commerce, healthcare, real estate, education. The more complex and research-heavy the decision, the bigger the effect; impulse purchases benefit the least.

The percentage of answers (within your monitored set of prompts) in which your brand appears, measured against competitors. Watch for the trap: if the prompt set includes prompts containing your brand name, the figure just measures your own name's echo, not the market — always ask for a brand vs. non-brand split.

Yes, more easily than in classic SEO: in many categories, citations are spread across dozens of domains and no one dominates. Models respond to concrete criteria (delivery, certification, price, specialization), where a niche site can beat a large generalist brand.

AI Overviews draws on Google's index and ranking systems — classic SEO remains the foundation, plus a direct-answer structure. ChatGPT combines the model's memory with its own web search, so your presence in the sources it reads also matters: press, reviews, communities.

A system that answers a question directly, synthesizing information from the language model and from web sources, instead of returning a list of links. Examples: ChatGPT, Perplexity, Gemini, Google AI Mode, Copilot.

Both, depending on the question. The model holds knowledge frozen at training time (up to a cutoff) and decides per question whether to trigger a web search; questions about prices, rankings, or recent events almost always trigger a search, while general questions often get answered from memory. There's no classic per-answer cache, but some platforms reuse recent search results for similar questions.

The date up to which the model's training data extends; anything you published after the cutoff doesn't exist in its memory and can only surface through web search. That's why a new or rebranded company needs to be strong in live sources (press, reviews, an indexable site) rather than simply waiting for the next model version.

The mechanism by which the model searches the web for relevant documents, builds its answer on them, and cites them. It's essential for AEO: it's the only channel through which you can influence answers quickly, without waiting for the model to be retrained.

Through live search: days to weeks, as soon as the page is indexed by whichever engine the platform uses (Google for Gemini and AI Overviews, Bing plus its own index for ChatGPT, its own index for Perplexity). Inside the model's own memory: only at future versions, which means months or longer.

Both: public sites end up in the training data (unless you block training crawlers), and for questions that trigger active search, the model reads the live version. In practice, your site needs to work for both: stable, clear content that's accessible to bots.

Each platform has a different model, a different cutoff, a different search engine, and different preferred sources. There's no single AI — visibility has to be built and measured platform by platform.

A central one: most answer engines lean on a classic search index — Google for its own products, Bing and proprietary indexes for ChatGPT and others. If you aren't indexable and visible in classic search, you'll be hard to find for AI too.

Generally not — crawlers see what a logged-out visitor sees. If your most valuable answers sit behind a wall, they effectively don't exist for answer engines; publish open versions of the content you want cited.

Unevenly: Googlebot renders JavaScript, but many AI bots (GPTBot, PerplexityBot, ClaudeBot) mostly take the raw HTML. Critical information — answers, prices, specifications — needs to exist in the initial HTML or be delivered via server-side rendering.

Anchoring the answer in verifiable, explicitly cited sources, instead of generating purely from the model's memory. Platforms are pushing harder toward grounding precisely to reduce hallucinations — which is why being a good, citable source has become the central goal of AEO.

Yes, frequently: language changes which sources get consulted (.ro sites versus international ones) and, as a result, which brands get recommended. For the Romanian market, monitor and optimize in Romanian — including the content you want cited.

Yes: small models know less from memory and rely more heavily on search, so web sources become even more important. Platforms often route simple questions to small models — one more reason to be strong in the search layer.

Yes — the entire conversation context feeds into the prompt: if the user has mentioned a tight budget, the model favors brands perceived as affordable, even without the user asking for that explicitly. That's why a brand needs to be clearly associated with its attributes (price, segment, values) in public sources: the model matches the profile of the request to the profile of the brand.

Yes, on the major platforms: ChatGPT has persistent memory and references conversation history (expanded through the 2025–2026 updates), and Gemini personalizes based on the Google account. OpenAI explicitly confirms that memory influences shopping recommendations — two users asking the same question can get different products.

Three causes: personalization (memory, location, custom instructions), non-determinism (the model samples, it doesn't recite), and search (different results at different moments). Shopping studies show brands appearing in over 60% of answers for some users and almost never for others, for the exact same question.

Not directly — memory is per-user and private; there's no lever by which a brand can write into another user's memory. What you can influence is the shared raw material: the web sources, reviews, and associations the model draws on for every user.

Yes: platforms use country and language (and sometimes city) to choose sources and filter options — local delivery, availability, .ro versus .de domains. For AEO, that means clear local signals: delivery pages, addresses, Google Business reviews, local press.

Typically: the current conversation, saved memory, conversation history, custom instructions, language, approximate location, and, on some platforms, account signals. Exactly what and how much varies by platform and by the user's privacy settings.

Yes, if you measure from your regular account — memory and history contaminate the results. Measure from memory-free sessions (temporary chat, clean accounts, API) and average across multiple runs; only then do the numbers become comparable over time.

No — conversational feedback stays in that user's memory. Models don't learn in real time from individual conversations; opinions only end up counting for everyone if they appear in public sources (reviews, forums) that models read.

Yes: without memory and history, the answer reflects only the model, the search, and the question. It's the closest thing to a neutral measurement — and exactly why real users with active memory may see something different from what your reports show.

It makes it probabilistic: you optimize the odds of being recommended to the right profiles, not a fixed position like in SEO. In practice you're aiming for consistency — appearing across as many relevant context variations as possible — not a universal number one spot.

There's no public trust score with a fixed timeline. Trust plays out on two layers: in live search, you can be cited as soon as you're indexed and relevant (days to weeks); in the model's memory, reputation settles through repeated appearances in established sources and only becomes visible in the next versions (months). Consistency over 3–6 months is the realistic horizon.

Not constantly — there's no permanent per-brand monitor. For every question that triggers search, the model re-consults fresh sources from the index; whatever has changed in them shows up in subsequent answers, not retroactively inside the model.

The combination: it answers a specific question directly and completely, has concrete data (figures, prices, criteria), is clearly structured (headings, tables), comes from an authoritative domain, and is recent. Models prefer pages from which they can extract an answer with no effort.

Yes, increasingly so: shopping platforms explicitly cite review aggregators, and Perplexity and AI Overviews lean almost entirely on external sources. A brand with no review ecosystem is practically invisible on engines that search for every question.

Yes: Wikipedia remains a strong entity anchor in training data, and Reddit and forums have grown massively as a source, partly through licensing deals with the major labs. Authentic community discussions about your brand are a genuine AEO asset.

Models don't reliably detect the paid-content label — what matters is the domain, the tone, and consistency with the rest of the sources. A factual advertorial on a domain already cited by AI in your category can work; an over-the-top laudatory one, contradicted by other sources, gets ignored or diluted.

Yes — for LLMs, the text is the signal, not the link: your brand's co-occurrence with a category and its attributes builds the association even without a hyperlink. The link remains useful for crawling and for classic SEO.

Models weight information confirmed by multiple independent sources: a claim that appears only on your own site is weak; the same claim echoed by press, reviews, and directories becomes the category's accepted truth. That's why serious AEO is ecosystem work, not just site work.

The main risk isn't a formal AI penalty, but inefficiency plus classic SEO risk: Google penalizes, and AI feeds on Google's index. Low-quality sources don't get cited anyway — you lose the money twice over.

Partly: public posts do show up in citations (Instagram and Facebook already appear in category reports), but video content without text is hard to leverage. Rich text descriptions, captions, and answer-style posts increase the odds of being used as a source.

The principles, yes: demonstrable experience, expertise, authority, and trust translate directly into the very signals models look for — real authors, first-party data, external citations. It isn't an identical algorithm, but E-E-A-T content is also citable content.

Yes, it's a real risk: each version has different training data, different weighting, and different search behavior, and studies show visible brand rotation between versions. The good news: brands with a solid presence in external sources rarely lose everything, because the live search layer stays in place.

Partly: new versions are trained on the web where you already appear, so your public footprint (press, reviews, Wikipedia, communities) carries over. What doesn't necessarily carry over is the exact behavior — which is why you re-measure after every major release.

Often: major versions every few months, with adjustments (routing, system prompts, search tools) happening virtually continuously, frequently unannounced. Treat AI visibility as a recurring series of measurements, not a one-off audit.

After major releases, yes — using the same set of prompts, so you can separate the model's effect from the effect of your own work. Monitoring tools do this automatically through daily or weekly runs.

Yes: post-training alignment changes the tone and caution of recommendations — for instance, a preference for consensus and well-documented options — which favors brands with a solid public footprint. One more argument for source consensus rather than one-off presence.

It can change overnight, because the candidate sources change: a different index, different partnerships. Diversification — being citable in Google's index, in Bing, and in licensed sources — is the only insurance policy.

Yes: model releases, system prompt changes, and the seasonality of sources (best-of-the-year lists, Black Friday) all produce waves of rotation. Publishing and updating content ahead of these windows is a tactical advantage.

For most consumer brands, marginally today — but they power third-party AI search engines, apps, and agents. The same public sources that make you visible in ChatGPT also propagate you through the open-source ecosystem.

There's no special vocabulary for AI, but standard schema.org is exactly what answer engines consume too: Product with Offer and price, Review and AggregateRating, Organization with sameAs, FAQPage. The rule: structured data must reflect visible content — messages hidden solely for AI don't work and erode trust.

For commerce: Product, Offer (price, currency, stock) and AggregateRating; for the brand: Organization with sameAs pointing to official profiles; for content: FAQPage, HowTo, Article with an author. Price and rating are the priority — exactly what AI cites in comparisons.

A proposed file that summarizes a site for AI agents. The current reality: Google has explicitly stated it doesn't use it for Search or AI Overviews, and no major lab has confirmed it as a signal in their search products; it mainly makes sense for technical documentation aimed at agents. It doesn't hurt, but it isn't the main lever.

If you want visibility, yes — blocking them removes you from training data and, on some platforms, from live answers too. It's a strategic decision (visibility versus control over your content), but for a commercial brand, blocking is almost always a net loss.

In your server logs or your CDN (Cloudflare has dedicated AI crawler reports): look for user agents like GPTBot, OAI-SearchBot, PerplexityBot, ClaudeBot, Google-Extended, then see which pages they read and how often. It's the most concrete sign of AI's interest in your site.

Indirectly: AI crawlers work with a limited time and rendering budget, and slow or heavy pages get read incompletely. On top of that, AI Overviews inherits Google's ranking, where speed matters. It isn't the decisive factor, but a slow site starts at a disadvantage.

Yes, if they're real questions — phrased the way people actually search — with direct answers in the first one to two sentences: exactly the format models extract most easily. Decorative FAQs with vague answers accomplish nothing.

Yes: models segment a page by headings, and an H2 that contains the exact question, followed by a concise answer, is the template with the highest citation rate. Creative but vague headings bury the content.

Yes, increasingly so: Google feeds its AI shopping experiences from Merchant Center, and OpenAI launched Instant Checkout on the Agentic Commerce Protocol (together with Stripe; Etsy and Shopify among the first partners), where products enter through dedicated feeds. For e-commerce, a clean feed becomes AEO infrastructure.

They remain the basic contract with every bot: the sitemap speeds up discovery of new pages, robots.txt decides who gets in — including AI bots, with separate directives for training and for search. Check that you're not accidentally blocking them with old rules.

A canonical about-brand page plus Organization schema with sameAs links to every official profile, consistent data (name, address, description) everywhere, and, where warranted, a Wikidata or Wikipedia entry. Entity confusion — a same-name brand, a rebrand — is a frequent cause of wrong answers.

Almost never: an indexable, fast site with clean HTML and structured data covers the technical side. What actually makes the difference is topical coverage — answering the questions buyers ask — and the ecosystem of external sources, not the technical platform.

Yes, and it's the advertorial's biggest lever: a direct answer to a specific question in the first paragraph, the question phrased as an H2, a comparison table with criteria and prices, concrete figures, the brand explicitly associated with its attributes. A brand-story-style advertorial with no answer structure is almost useless for AI.

The title = the real question; the first two sentences = the answer; then criteria, a comparison table, real prices or ranges, a short FAQ, author, and date. The practical test: if the answer can be copied straight from the first lines, the model can extract it too.

For questions like how do I choose or which one is better, models cite decision guides built around criteria, not product grids. A category page sells after the click; a guide gets you into the answer. You need both, linked to each other.

You can, and it works if it's honest: transparent criteria, real competitors included, and yourself positioned where you have actual arguments. Models synthesize existing lists — if the only rankings in the market belong to others, their ranking becomes the AI's ranking.

Length isn't the criterion — completeness on a single question is. A 600-word answer that covers the question fully beats a 4,000-word pillar that only touches on it in passing; group related questions into separate, interlinked pages.

Yes, especially for live search: models prefer recent sources for commercial questions, and last year's rankings lose citations. Real updates — prices, new models, a visible date — not just changing the year in the title.

Not because it's generated, but if it's generic: models cite new, specific, verifiable information — exactly what unfiltered AI text lacks. AI as a drafting tool built on first-party data works; AI as a filler factory stays invisible.

They're among the strongest assets: unique figures — studies, surveys, internal benchmarks — force the model to cite you as the origin, rather than just paraphrase you. One good study a year can bring more citations than dozens of articles.

Yes, if done factually: X versus Y questions are common, and if you don't answer them, third parties will, or the model will improvise. Be fair about the competitor's strengths — the comparison's credibility determines whether it gets cited.

Every question a customer asks from awareness to decision: definitions, selection criteria, comparisons, prices, common mistakes, post-purchase maintenance. In practice, the prompt map from your monitoring becomes your editorial plan.

Through their text: transcripts, rich descriptions, and supporting pages are indexable and citable; YouTube is frequently cited in AI Overviews. Publish the transcript of every episode or video as a structured page.

There's no stored ranking: the top list is generated for every single question, from the model's memory plus whatever sources are found at that moment plus the user's context. Other users get similar rankings to the extent that they draw on the same sources — which is why the sources (published lists, reviews) are the real ranking to influence.

Not as an explicit list — there are statistical associations: how tightly the brand is linked to the category and to its attributes in the data the model has seen. In practice it works like a probability of being evoked, one you raise through consistent presence in sources, not a spot you conquer once and for all.

Models generate probabilistically (they sample), and web search returns slightly different result sets from one run to the next; personalization stacks on top of all that. That's why serious measurement averages across multiple runs — a single run is an anecdote, not data.

Mainly by synthesizing existing lists: editorial rankings, comparison sites, aggregated reviews, community discussions. If you appear consistently in these materials with the same attributes, you make it into the synthesized top list; if you don't exist in them, the model has nowhere to pull you from.

You identify the sources the model cites for that prompt (Perplexity and AI Overviews display them), then work directly on those: PR toward those publications, presence in comparison sites, an honest list of your own, reviews. You target the answer's sources, not the answer itself.

Yes: the first recommendation captures a disproportionate share of attention, just like in the SERP. But in AI, the difference between positions 2 through 5 is smaller than the difference between being in the answer and not being in it at all — consistent presence is the priority.

Ad formats do exist and are expanding (Perplexity was the pioneer, others are testing), but they're labeled as sponsored and kept separate from the organic recommendation. The actual recommendation can't be bought on any major platform — it's earned through sources.

Fixed prompts, run identically, from clean environments (API or temporary chat), across every relevant platform, averaged over multiple runs, at a regular cadence. That's exactly what AI monitoring platforms automate.

Yes: if the dominant sources about you are negative — bad reviews, press scandals — the model synthesizes the warnings too. Monitoring the sentiment of mentions is just as important as the mention rate.

Substantially: each language draws on a different body of sources, and therefore reaches a different consensus. A brand can be the leading answer in Romanian and invisible in English — treat each market as a separate AEO project.

Follow your audience: AI Overviews has the largest volume (it sits on top of Google searches), ChatGPT has the widest conversational adoption, and Perplexity attracts research-oriented users. In practice, an SEO foundation plus answer-ready structure serves AI Overviews, while an ecosystem of external sources serves ChatGPT and Perplexity — you end up covering them in parallel more than it seems.

It's built retrieval-first: it searches for every question and cites sources as a matter of course. The consequence: you can't be visible on Perplexity without external sources — reviews, press, listings. Your own site isn't enough.

Google triggers it selectively, on informational or complex queries, and holds it back on simple transactional ones, on sensitive topics, or wherever classic results are already sufficient. AI Overviews not showing up on a keyword isn't your failure — you monitor only where it actually triggers.

AI Overviews = a summary sitting above classic results; AI Mode = a full conversational search experience, with follow-up questions and longer synthesis. AI Mode draws on more sources per answer and cuts into clicks more aggressively.

Yes, more than their apparent weight suggests: Bing's index also feeds other answer engines, and Copilot is baked into the Microsoft ecosystem (Windows, Office, Edge) — especially relevant for B2B. Check your indexing in Bing Webmaster Tools; it's frequently neglected.

Through the data it can read: commerce feeds and protocols (Agentic Commerce Protocol / Instant Checkout, built with Stripe, with Etsy and Shopify among the first partners), plus correct Product schema and aggregated reviews. Agentic commerce is still early, but it's moving fast — worth a pilot if you run e-commerce.

Software that searches, compares, and completes purchases on behalf of the user — the direction known as agentic commerce. Preparation means impeccable product data (price, stock, returns, delivery, in structured format), frictionless checkout, and clear policies — the agent picks whatever it can process without ambiguity.

As a secondary priority: their adoption for purchase decisions trails ChatGPT and Google, but it's growing in niches (Claude for professional and B2B use, Grok for real-time on X, Meta AI for social and mobile). The rule: monitor wherever your audience actually asks; strong sources cover you everywhere regardless.

Yes, as they move onto LLMs: a voice answer is the absolute number-one position — a single answer, zero list. The same principles (a direct answer, a clear entity, local data) serve voice too.

Yes: OpenAI has GPTBot (training) and OAI-SearchBot (search), Google has Googlebot (search) and Google-Extended (AI training), Anthropic has ClaudeBot, Perplexity has PerplexityBot. For commercial visibility: allow all the search bots; decide training access separately, knowing that blocking it reduces your footprint in future models.

There's no official equivalent — platforms don't report back to brands. Visibility has to be measured actively: a fixed set of prompts run recurrently on each platform, plus citation analysis and AI referral traffic in your analytics.

The core set: mention rate on purchase-intent prompts (not just brand ones), average position within the answer, mention sentiment, the number of domain citations, share of voice against competitors, plus traffic and conversions from AI referrals. A brand versus non-brand split is mandatory so you don't fool yourself.

You segment referrals: chatgpt.com, perplexity.ai, gemini.google.com, copilot.microsoft.com, and the rest. Careful: some AI traffic arrives labeled as direct (dark traffic), so the numbers you see are an underestimate.

Studies from 2025–2026 converge on the same conclusion: on average 4–5 times better (Adobe reported AI traffic converting 42% above average in retail in March 2026), with wide variation by industry. The user arrives pre-qualified, since the AI has already done their research for them: lower volume, higher value per visit.

Order of magnitude: 30–50 prompts (at least 70% non-brand, phrased the way a real customer would ask), each run across all relevant platforms, repeated at a fixed cadence. Conclusions should be drawn from multi-week trends, not a single day's snapshot.

Weekly for trend series, monthly for management reporting, plus ad-hoc measurements after major model releases or your own campaigns. Daily only if a tool does it automatically — manually, it isn't worth it.

The market has matured: Profound (enterprise, starting around USD 499/month), Peec AI (mid-market), Otterly (entry-level, starting around USD 29/month), plus modules inside SEO suites like Semrush and agencies with proprietary monitoring. To get started, a one-off audit plus monthly monitoring is enough; a dedicated platform earns its keep at scale.

You classify each mention — positive, neutral, negative — plus the context: recommended, merely compared, or warned against. Manually at low volume, with an LLM at high volume. Sentiment separates showing up often from showing up well; a brand mentioned often but negatively has a problem, not a success.

Imperfectly, like all brand marketing: combining AI referral traffic with its conversions, a where did you hear about us question at checkout, dedicated codes, and correlating the mention trend with direct traffic and brand searches. There's no complete click-path — anyone promising perfect attribution is overselling.

Not the ones from the physical market, but the ones actually winning the answers: you run the prompts and see who shows up and who gets cited. The resulting list is almost always surprising — often online players you weren't tracking at all.

Common models: a one-off audit, a monthly subscription (monitoring, content, digital PR), or a phased project; serious programs start around the level of an average SEO retainer and scale up with content volume. The expensive component is producing citable content, not monitoring.

The first movements in the search layer (Perplexity, AI Overviews, citations) show up 30–90 days after publishing the right content; consolidation in the models' memory takes 6–12 months. Anyone promising domination in two weeks is selling something else.

At minimum: a per-platform visibility audit with a brand/non-brand split, a documented prompt set, a fair competitive benchmark, a prioritized content and technical plan, a monthly report with the same metrics comparable over time, and access to the raw data. Without verifiable raw data, a report is just a story.

Ask about: the measurement methodology (how many runs, from what environments), the brand/non-brand split, how they select competitors, and examples of content that has actually won citations. Red flags: guarantees of position, claimed partnerships with OpenAI or Google invoked to justify organic recommendations, reports built only on brand prompts.

No — organic recommendations have no paid or guaranteed channel on any major platform, and non-determinism makes any position guarantee dishonest. Only the actions taken and their measurement can be guaranteed, not the models' behavior.

A hybrid works most often: strategy and expert content stay in-house (you know your product best), while measurement methodology, the technical side, and digital PR get outsourced. Going fully in-house requires a dedicated person who keeps up with the platforms' monthly changes.

The order with the best cost-to-impact ratio: one, an audit on 10–15 real purchase-intent prompts; two, fix whatever is technically blocking you (crawlers, schema); three, 5–7 decision guides on the questions where you have 0% visibility; four, Google Business reviews. Only after that, PR and paid monitoring.

It isn't a new silo: SEO ensures indexability, PR builds the sources that get cited, and social feeds the communities models read. AEO adds a layer of orchestration — which questions we target, in which sources, in what format. Budget-wise it's a reallocation, not a parallel channel.

The category settles without you: the competitors occupying the answers today become the consensus the models learn and repeat, and dislodging them later costs multiples of what it would have cost to claim an open window. In categories with no AI leader yet, the first-mover advantage is real.

A market benchmark: 10–30% of your search and content budget, more if your audience is early-adopter or the purchase decision requires heavy research. Test for one to two quarters, measure mentions, AI traffic, and conversions, then scale whatever proves out.

Three lines: search volume is migrating to AI (Gartner projects a 25% drop in classic search by 2026); AI traffic converts roughly 4–5 times better; the cost of claiming a category rises with every quarter of delay. Plus your own measurement: today we appear in X% of the answers to our customers' questions.

Three directions: correct the sources it feeds on (your own site stated explicitly, press, reviews, Wikipedia), report it through the platform's feedback mechanisms (aggregated feedback carries weight), and publish official pages that directly address the specific confusion. Live search picks up corrections within weeks; the model's memory only at the next versions.

There's practically no way to delete something from a model that's already been trained; correction happens in layers: correct live sources dominate search-based answers, while personal data (as opposed to brand data) has separate legal channels through GDPR. The realistic strategy is to make the truth more visible than the error.

Yes: models synthesize warnings too — for example, customers complain about delivery — and a recurring negative theme in reviews becomes part of the brand's profile. Fixing the actual underlying issue, plus a volume of new positive reviews, dilutes the signal; cosmetic fixes don't.

There are attempts — text hidden for bots, instructions injected into pages, fake reviews, content farms — and they die out quickly: platforms actively filter for them, and the collateral penalty comes from classic SEO. The short window of gain doesn't offset the risk of being excluded from sources.

Instructions hidden inside web pages that try to manipulate the model when it reads the page — for example, recommend this brand. Platforms treat these as a security attack; for a brand, getting caught means losing trust — the very asset AEO is meant to build.

It happens — changes to crawlers, to commerce, to formats — which is why the durable assets are the ones no single platform controls: your content, your reviews, your press, your community. Diversifying across platforms plus a solid source foundation is your insurance policy.

It's a genuine trade-off: blocking (Google-Extended, GPTBot) protects your content but reduces your footprint in future models. It can make sense for publishers; for a brand that wants to be recommended, generally not.

Yes, it's partly the new normal: many answers get consumed without a click (zero-click), and value is migrating from traffic to presence within the answer. Track these together: mentions, AI traffic (lower volume, but more qualified), and conversions — not just total sessions.

Users' conversations belong to the platforms, not to you — you don't get their search data, not even the equivalent of search terms. Normal GDPR applies to your own site; what's new is that the customer's research moves into a space you can't directly observe.