Frequently Asked Questions

GEO stands for Generative Engine Optimization. SEO aims to rank a page on a results page; GEO aims to be cited inside an answer that an AI model has already written. The practical difference is that a model needs extractable, attributable facts rather than persuasive copy. Crawlability, structured data and clarity still matter, but attribution and entity consistency become the deciding factors.

Use a fixed question set and measure mention rate, not traffic. Pick 30 to 60 questions a real buyer would ask, run them on a fixed schedule, and record whether you are mentioned, where in the answer you appear, and which source was cited. Judge a trend rather than a single snapshot, because model answers drift for reasons unrelated to your work.

Write both, for different jobs. Short, self-contained answers that survive being lifted out of a page suit extraction; long-form explanations establish authority and supply the context a model needs to trust a claim. If a page has only length, it buries the fact. If it has only punchy slogans, it gives a model nothing verifiable to cite.

It is a proposed convention, not an enforced standard. No major AI platform currently requires or specifically rewards it. It is cheap to publish and may be adopted, but it should never replace correct robots.txt rules, which are what actually control whether a crawler can fetch your pages at all.

Technical fixes such as crawlability and schema often show up within days because they only require a re-crawl. Content and authority changes take longer, typically several weeks, because they depend on your pages being retrieved and reconsidered. Establish a baseline question set before you start; without one you cannot tell improvement from normal fluctuation.

Not directly. Citations frequently answer the question without a click, so traffic is a poor proxy for AI visibility. Measure mention rate and position directly, and if you need to attribute revenue, set up trackable links or check whether branded search volume moves after you start being cited.

Cost comes in three parts: a one-off diagnosis covering crawlability, structure, content and baseline measurement; ongoing content and structural work; and repeated measurement. What moves the number is site size, how many AI platforms you need to cover, and how much restructuring the content needs — not brand awareness. For most mid-sized sites, running a diagnosis before committing to a longer engagement is the sensible order: it tells you what is actually broken before anyone quotes a retainer. YupuGEO offers a free diagnosis and free query tools, so the starting point costs nothing to establish.

Three things must be fixed in writing before work starts, because without them no comparison means anything: the question set (thirty to fifty questions covering informational, comparison and decision intent), the platform list, and how questions are asked (fixed time window and phrasing). Pass criteria differ by metric type. Fundamentals — crawlers allowed, schema valid, pages fetchable — should pass completely. Citation coverage means some pages you control get cited, not site-wide coverage. Answer accuracy cannot be traded for volume. Business outcomes are observed over a longer window and are not a basis for accepting a single delivery.

Several signals need no expertise to read. Promised rankings or guaranteed citations — model answers cannot be promised. Screenshots without the raw responses, so nothing can be reproduced. No answer on measurement: how many questions, which platforms, what time window. Conflating being mentioned with being recommended. Treating domestic and international engines as interchangeable. A credible engagement delivers before-and-after results on the same question set, the full AI responses, a fix list naming specific pages and positions, and a clear line between what was verified and what is still open.

In-house GEO needs three capabilities that rarely sit together: engineering (robots, schema, rendering, crawl logs), content (turning business knowledge into quotable structure), and data (fixed question sets and before-and-after comparison). The barrier is not budget but having all three. Most teams do not need to choose between fully in-house and fully outsourced: buy the diagnosis and the baseline, then execute internally. Diagnosis happens once; execution happens continuously. That capability has to be built rather than bought.

Triage in this order. First, check the fundamentals: is robots.txt blocking the major AI crawlers, can target pages be fetched at all, and do they return real content rather than a login wall or a JavaScript-only page. Second, ask whether the content can be quoted — retrieval systems need a part of the page they can assemble into an answer, and one block of promotional prose offers nothing. Third, check whether you are asking the question buyers actually ask. A large share of cases that look like a failed method turn out to be a gap in one of the first two layers.

Model output is probabilistic. The same question can return different sources at different times with no change on your side. Without a baseline you cannot distinguish your changes from sampling noise, and you cannot attribute movement to anything. The method is simple: fix thirty to fifty questions, fix the platforms, fix the time window, and record for each whether the brand appeared, which pages were cited, which competitors appeared, and the complete AI response. That dataset is what every later comparison is made against.

GEO has the most leverage where decisions are long and heavily discussed: professional services, healthcare, legal, education and B2B manufacturing, where buyers ask for a recommendation before they ask about a vendor. It also suits high ticket sizes and long research cycles, and teams that already have content assets to restructure. It has the least leverage in impulse purchases where nobody researches, in categories so commoditised that an AI answer cannot distinguish brands, and on teams with no plan to produce content at all.

Major AI crawlers respect robots.txt. Blocking them does not hide anything from anyone — it removes your content from the systems you are trying to be visible in, and nothing else. The more common accidental block is a path pattern written for search crawlers that also catches the AI ones, or a CDN rule that returns a challenge page to unknown user agents. If AI cannot fetch your pages, no amount of content work will change anything, so this is the first thing to verify.

It has to be readable, and that is not always the same as being client-rendered. A page that assembles itself only in the browser may return little or nothing to a crawler that does not execute JavaScript. Server-side rendering or pre-rendering removes the uncertainty. This matters more for AI visibility than it does for traditional search, because some retrieval systems rely on the raw HTML response rather than a full browser environment.

Structured data does not directly cause citation, but it removes ambiguity about what a page contains. Schema that states the organisation, its products, prices and FAQ content lets a retrieval system interpret the page without inferring it. The practical effect is on pages where the same information appears in prose: machine-readable markup reduces the chance of it being misread or discarded. It is one of the cheaper improvements available, but it is not a substitute for content that answers a question directly.

Retrieval systems assemble part of a page into an answer, so what matters is whether a usable piece can be lifted out cleanly. The conclusion needs to appear in the opening lines rather than after a preamble. Each second-level heading should work as a standalone claim. Facts need to be specific enough to check — numbers, dates, versions, conditions. Tables and lists tend to survive extraction better than flowing paragraphs, because their boundaries are already explicit.

From conversations that already happened, not from a brainstorm. Recorded sales and support calls are the highest-quality source because they contain the exact phrasing buyers used. The queries your own site search receives come next. Questions in competitor reviews are third. The usual failure is a team writing questions that sound reasonable and resemble nothing buyers ask, which produces data that cannot be acted on and content nobody searches for.

It is a partial result with a clear next step. Being named means the brand entity is recognised. Not being cited means the content is still not quotable — retrieval had nothing clean enough to point at. What to fix is usually concentrated: the pages that answer questions in your category should lead with a conclusion, state checkable facts, and have headings that work as standalone claims. The brand recognition will not regress while you work on that.

It is cheap and occasionally useful, but treat it as a tidy-up rather than a lever. It helps only if the content it points at is already quotable, and adoption across engines is uneven. Where it does earn its place is as a maintained summary that makes your canonical facts easy to locate — pricing, leadership, product scope — and reduces the chance of an engine assembling a stale version from somewhere else. If you create one, keep it current; a stale summary is worse than none.

In this context it describes content published to influence how a retrieval-based answer describes a brand, rather than a technical breach of a model. The pages involved are ordinary web pages: comparison posts, review aggregators, or FAQ entries that state facts about your category with your brand as the negative example. Because they are ordinary pages, the defence is ordinary too. Publish a more specific, more current and more fetchable version of the same fact on a page you control, and confirm that AI crawlers can actually reach it. Sources reviewed on 2026-10-20. This describes a method with verifiable boundaries and does not claim that any platform will change how it describes a brand.

Two patterns distinguish it. Direct injection leaves phrasing in the answer with no traceable source page, which suggests something was retrieved that you cannot locate. Entity conflict is easier to trace: the claim is published on a page you can find, and it competes with your own version of the same fact. To separate the two, search the exact wording that appeared in the answer, and record any third-party URLs cited for your brand questions across rounds. Neither pattern proves intent, so treat a signal as grounds for review rather than as proof. Sources reviewed on 2026-10-20.

Three paths, and the choice depends on the type of claim. Publish the fact: state the specific correction, dated, on a page that AI crawlers can fetch. Dispute the source: where the claim is demonstrably false and a platform operates a claim process, though this is slow. Leave it: where correcting would require making a claim you cannot support, since an unsupported correction is retrievable evidence against you later. There is no mechanism for having a model provider remove a page from an answer. What can change is which evidence is stronger and more current. Sources reviewed on 2026-10-20.

Four sources are distinguishable and each calls for a different response. Stale retrieval: the answer describes a past version, so recency is the lever. Entity conflation: your facts are present but attached to the wrong entity, so consistency of your own attributes is the lever. Unsupported inference: the model filled a gap from adjacent facts, so state the fact explicitly somewhere quotable. Third-party sourced: a cited page asserted it, which you cannot correct directly. Classifying before responding is what makes the correction worth doing. Sources reviewed on 2026-10-20.

Because the persistence of a wrong answer comes from the retrievable evidence rather than from the model. A model update changes how existing evidence is assembled; it does not determine which source is correct. If a third-party page is the clearest source on a fact, it stays retrievable indefinitely. There is also no notification path — nobody is told that an answer about your brand became wrong, and there is no ticket unless a content policy was breached. A correction that depends on the platform noticing has no owner and produces no record. Sources reviewed on 2026-10-20.

Design the test before making the change. Record the exact question string, platform, timestamp and full response text of the erroneous answer, then repeat that question after the corrected page has had time to be crawled. Two cautions follow. A single re-test proving nothing is not evidence of failure, because sampling varies run to run, so repeat across a window. And a re-test that shows improvement may reflect sampling coincidence rather than the correction, which is why several rounds are needed. Keep the fixed question in the standing set so a regression stays visible. Sources reviewed on 2026-10-20.

They are three states, and collapsing them into a single visibility rate discards the only information that distinguishes commercial value. Mentioned: the brand appears with no link, which indicates entity recognition. Cited: a page you control appears as a source, which indicates the content is quotable and was retrieved. Recommended: the answer selects the brand, which indicates comparative advantage in the evidence available. Only the third is close to a decision, and the three call for different optimisation work. Sources reviewed on 2026-10-20.

Confirm first that the AI domains appear as referrers rather than being flattened into direct traffic, which is what consent tooling usually does. Then add a tagged landing path so the source survives any internal redirect, and capture referrer, landing page and first-touch source together. One structural limitation matters more than the tooling: an answer can cite your page and the reader can act on it without ever visiting, so no analytics configuration captures that interest. Treat referrals as a measurable subset of AI-driven demand, not as the whole of it. Sources reviewed on 2026-10-20.

Neither last-click nor first-click is defensible here, because AI referrals often arrive immediately before an enquiry rather than causing one. Accept that AI cannot be fully attributed and define what each signal contributes instead. Count citations as exposure, tracked in the question set rather than in analytics. Track AI-referred sessions as a floor on real volume. Use enquiries as the business number on a quarterly window, kept separate from weekly visibility reporting. Report a count rather than a rate, because any rate built on incomplete referral data has a denominator nobody can verify. Sources reviewed on 2026-10-20.

Order by how much of the gap each piece closes and what happens if it stays open. Factual defects come first: wrong information is retrieved and repeated, so it costs more than absence. Then questions where nobody is cited, since there is no competition and the position does not close by itself. Then questions where a competitor holds the citation, which depends on their page ageing or yours improving. Already-cited pages come last as maintenance. Do not rank by traffic volume — AI visibility and organic traffic are different distributions. Sources reviewed on 2026-10-20.

The fields whose absence makes priority undecidable. The fixed question verbatim as asked during measurement. Platform and date of the observation. Current state — mentioned, cited, recommended or absent, which are three different starting positions. Who is cited today, with URLs, so the competitive claim is specific. Which of your pages answer it today. Estimated effort and owner. And the acceptance check: what will be re-run, when, and what counts as a pass. A request carrying any two of these is a wish, not a ticket. Sources reviewed on 2026-10-20.

Specify the exact question rather than the topic, the direct answer in two to four sentences at the top, key facts as a list or table, one h2 per sub-question written to stand alone as a claim, and the objection a reader would raise. Acceptance should check four things: the page is fetchable with the answer present in raw HTML, the structure puts the conclusion first, every fact matches the source of record, and the same question is re-tested after enough time for crawl and index updates. Citation itself should not be an acceptance criterion, since retrieval is not a publisher's to command. Sources reviewed on 2026-10-20.

One ledger entry per claim, not per page, and the fields are short: the claim in the exact wording used on the page, the source publisher and last-updated date, the retrieved-on date and URL, the population and methodology as the source describes them, what the source does not claim, who checked it and when it was last re-verified. Re-verify on the page review cycle rather than on a separate schedule. Where a claim can no longer be supported, remove it rather than softening it — a softened claim still gets extracted as an assertion. Sources reviewed on 2026-10-20.

Two failure modes. The outdated table: accurate when published, wrong now, still live, and usually the most retrievable page on the site because comparison questions have clear answers. The unqualified superlative: written as marketing, read as a factual assertion once extracted without its surrounding context. Mitigations are practical. Date the comparison on the page itself rather than only in a changelog. Link each competitor row to a source. Scope superlatives explicitly. Treat an aging comparison as a defect rather than as a low-priority refresh. Sources reviewed on 2026-10-20.

Draw the line between figures from your own records and figures attributed to customers. The first is auditable: state the measurement method, the period and the population. The second needs a permission trail, agreed wording and an expiry date, and without all three it is an unsupported claim that will be quoted back. State the sample size for any aggregate — a percentage of ten is not a percentage of ten thousand. Put the condition and the date inside the sentence carrying the number, because a qualifier in the following paragraph is dropped during extraction. No projection should be presented as a result. Sources reviewed on 2026-10-20.

Four variables account for almost all of it. Model version: the underlying model is replaced and phrasing and source selection shift. Live web access: the same question answers differently depending on whether the system retrieved anything that day. Question phrasing: small rewording changes which sources are retrieved, sometimes more than any site-side change. Sampling randomness: the same question on the same day returns different answers with no identifiable cause. Only phrasing is under your control, so that is the one to fix first in any comparison. Sources reviewed on 2026-10-20.

Three to five repetitions per round, spread across the window rather than run back to back. The spread across those runs is the noise floor for that question, and any later movement smaller than it is unexplained. A question whose answers swing between three and six cited sources is not a usable measurement instrument whatever its baseline says, so questions differ in how much weight they can carry. Record model version where the interface exposes it, note whether anything was cited at all since live retrieval is the strongest predictor of volatility, and store every response verbatim. Sources reviewed on 2026-10-20.

Check three things before crediting anything. Does the movement exceed the spread measured for that question in the baseline round. Does it appear on platforms where nothing was changed. Does it persist across several rounds rather than appearing once. Movement confined to one platform and one day is usually noise or a model update. Movement across platforms, persisting across rounds and exceeding the recorded spread is the only category worth reporting as attributable. Report unexplained movement as unexplained — a gap in the record is easier to correct later than a guess. Sources reviewed on 2026-10-20.

Because retrieval has no basis for choosing between conflicting versions and tends to take the one most prominent for the language the question is asked in. Two effects make the outcome fairly deterministic. Specificity beats generalism: a page stating a precise number is more extractable than one stating a range, so precision decides the conflict whether or not it is current. Language alignment matters more than expected, so an out-of-date English page can shadow a current one for English questions. A version that was translated and then abandoned stays retrievable; nothing retires it automatically. Sources reviewed on 2026-10-20.

Organisational discipline first, tooling second. Keep one document as the source of record for founding year, legal name, headquarters and product scope. Give each attribute a named owner with a change process that requires updating every language version. Repeat identical structured data in each version, including attributes that legitimately differ by market. Declare hreflang reciprocally so the relationship is explicit. Where facts genuinely differ by region, state the difference in the text of each version rather than omitting it, because variation that exists only in an internal spreadsheet reads as a contradiction. Sources reviewed on 2026-10-20.