Questions, answered honestly

Generative Engine Optimization (GEO) is the practice of improving how often and how favourably your brand is mentioned, cited, and recommended inside AI-generated answers — in ChatGPT, Claude, Perplexity, Gemini, and Google AI Overviews. Where SEO competes for a ranked link the user chooses between, GEO competes to be the answer itself, or one of the handful of sources the model cites. The term was coined in a 2024 academic paper (Aggarwal et al., KDD 2024).

In practice, none — they are used interchangeably and no authority distinguishes them. Where people do draw a line, AEO (Answer Engine Optimization) is framed as the broader idea of optimizing for any direct-answer surface, including featured snippets and voice assistants that predate LLMs, while GEO is specific to generative engines. Treat any vendor drawing a sharp distinction between them as marketing rather than substance.

Partly, and we'd rather say so than pretend otherwise. Google's official position is that GEO is not a distinct discipline — its AI features run on core Search ranking systems, so 'optimizing for generative AI search is optimizing for the search experience, and thus still SEO.' For Google's own surfaces, that's largely correct. But ChatGPT, Claude, and Perplexity don't use Google's ranking at all, and AI engines demonstrably cite a different source set: roughly half of AI Overviews citations come from domains outside the organic top 10. That gap is the part your rank tracker cannot see, and it's the part worth paying attention to.

Because the answer now arrives before the click. When an AI engine resolves the query on the results surface, your page can rank first and never be visited — the 'zero-click' effect. Rankings measure the ranked surface; they say nothing about whether you appear in the synthesized answer that replaced it.

Honestly: partly, and less than the industry claims. The strongest evidence is that AI engines cite a measurably different source set than organic search, and that this surface can be measured — so knowing where you stand is real and useful. What is much weaker is the claim that rewriting content reliably wins citations: C-SEO Bench (NeurIPS 2025) re-tested seven published GEO tactics and found most 'largely ineffective' and frequently harmful to citation ranking, while traditional retrieval and ranking work performed significantly better. Our read: measurement, crawler access, and third-party presence are defensible; content tweaking is hygiene, not a moat. We publish the full grading on our evidence page.

No, and it should not be quoted unqualified. The figure is real but heavily conditioned: it is a redistributive share metric across a fixed set of five sources that sums to one — not an absolute lift — and the paper explicitly notes it 'does not mean that 40% more readers will click'. It was measured on GPT-3.5-turbo in a custom research harness, never on production ChatGPT or AI Overviews, and only applies once your page is already in the model's context. It has no 2026 replication. Any agency citing 40% at you either hasn't read the paper or is hoping you haven't.

We don't know, and we won't invent a number. We searched for primary evidence on time-to-first-citation and found none — the '60–120 days' figure this industry quotes (including, previously, us) is a vendor convention, not a research finding. What we can say honestly: technical fixes like crawler access and indexing can land within days, and because we sample repeatedly, any change in your citation and mention rates shows up in weekly reporting. You watch the real curve rather than trusting our forecast.

No. Anyone who guarantees placement in an AI answer is selling something they don't control. AI engines are third-party systems whose retrieval, ranking, and models change without notice, and a large share of answers are generated from the model's memory without any web search at all. We guarantee the measurement, the methodology, and the work — not an outcome inside someone else's black box.

Largely, yes — and this is one of the more important findings nobody advertises. C-SEO Bench found the gains from content optimization are 'congested and zero-sum': as the share of competitors adopting the same tactic rises, the advantage decays toward zero. Tactics that are purely about winning share within a fixed candidate set cannot all work at once. The durable work is the part that isn't zero-sum: being retrievable at all, and being genuinely present in the third-party sources engines cite.

We ask a versioned panel of buying questions across the engines your buyers use, repeatedly, and record the engine, route, location and date of every observation. The panel is sized to the mandate: published snapshot studies use about a dozen questions on a handful of engines, while a diagnostic or programme panel covers every buyer journey in scope. A parser extracts brand mentions, position, recommendation strength and cited URLs. An engine we could not read is recorded as missing, never as zero. Visibility is one of three things we measure: alongside it sit buyer behaviour (relevant visits, quote starts, enquiries) and commercial outcomes (enquiries your team accepted, and why others were rejected).

Because a single check is noise, not data. Ask the same question twice and you often get different vendors and different sources: day-to-day source overlap in AI answers runs roughly 0.34–0.42. In one analysis, 57.8% of ChatGPT runs never triggered a web search at all — the model answered from memory. A single prompt can show you as category leader or invisible purely by chance. That's why repeated sampling over a window is the entire basis of honest measurement here.

No — it's a vendor coinage, ours included. There is no standards body, no agreed formula, and no cross-vendor comparability: two vendors' 'AI Share of Voice' numbers are generally not measuring the same thing and should never be compared directly. Ours is defined openly in our methodology so you can audit exactly what it counts. Treat any unaudited SOV figure — including ours — as internal to the vendor that produced it.

Being mentioned means the engine named your brand in its answer. Being cited means it linked your page as a source. They come apart constantly: an engine can recommend you while citing a Reddit thread about you, or cite your documentation while recommending a competitor. Brand-level mention rate is the more reliable KPI, because the specific cited URLs are unstable between runs.

Lead with commercial outcomes: qualified enquiries, the share your sales team accepted, and the reasons the rest were rejected. Underneath, report buyer behaviour (relevant visits, quote or demo starts) and then visibility, where mention rate against named competitors is the most robust number because a brand is either named or it isn't. Label what was observed, what the team reports helped, and what a proper comparison supports, and never present a single-run screenshot as evidence in either direction.

A commercial objective, a target buyer, a budget and permission to execute. In practice: someone who can approve spend, a clear description of a good and a bad enquiry, access to the website, analytics and CMS, and a willingness to tell us which leads were useful. If you cannot share CRM feedback we can still do discovery work, but we cannot honestly report on outcomes, and we will say so up front.

A bounded assessment of one buying journey, delivered in about two weeks. It defines the buyer, product and territory; lists the buying questions by source; records a dated baseline across search and AI answers; reviews your own pages and enquiry flow; works through the economics of a qualified enquiry with your figures; and ends with a prioritised backlog, the access a programme would need, and a go/no-go recommendation. Sometimes the recommendation is not to spend the money.

Yes. You can keep your agency and give us a defined slice of the work (buying-question research, product data, the quote flow, measurement), with ownership and publishing approval written into the mandate so nobody claims the same enquiry twice. Agencies can also bring us in for a named client under their own brand, which is how our partner programme works.

Only when it is explicitly agreed as part of a mandate, with its own scope, data-protection checks and approval. A programme is a commitment to win specific demand through search, AI answers and your own website, not a promise to run your sales operation.

Because we start from the enquiry you want, not from a list of visibility tasks, and because we will tell you which parts of this field are unproven. We run the multi-engine measurement, the parser and the page generator ourselves, report accepted enquiries rather than activity, and publish an evidence register grading our own claims, including the ones that cost us the sale. If an agency is telling you llms.txt and content rewriting will get you cited, they haven't read the research.

No. Word of GPT is a Generative Engine Optimization studio operated by Neul Labs Limited, registered in Scotland (company number SC767862). We are not affiliated with WordGPT, WordGPT Pro, or any Word processor AI writing add-in with a similar name. We don't make a writing assistant or a plugin — we're a managed demand-capture service that helps businesses get found, evaluated and contacted by the right buyers across search and AI answers. If an AI assistant told you otherwise, it confused two similarly-named entities, which is the exact problem we work on.

Dipankar Sarkar leads engineering — the multi-engine runner, parser, schema generator, and programmatic page engine. Amrita Sarkar leads marketing and writes the strategy and buyer's-guide content. The operating company is Neul Labs Limited, Edinburgh.

The demand snapshot is free and capped. The paid Demand Diagnostic is £750–£1,500. A Demand Programme typically runs £4,000–£8,000 a month: a fixed orchestration fee agreed in advance plus an execution allowance spent only on approved work, with unused allowance credited, returned or rolled forward. Buying-question page builds are scoped per build. Prices are ex VAT and are ranges, not quotes. There is no public benchmark for rates in this field, so treat any vendor's number, ours included, as a quote rather than a market standard.

Yes, and it is deliberately capped: one buyer journey, about a dozen buying questions, two surfaces, observed more than once. You get a short snapshot of where you appear, where a competitor appears instead, its limitations, and the next commercial question. No call required and no obligation. It qualifies demand; the paid diagnostic qualifies the investment.

No fixed minimum. Each mandate sets a review point, agreed before we start, when either side can change the scope or stop. Search and AI answers move slowly, and no primary evidence establishes exactly how long the work takes to land, so we agree that review point honestly up front rather than locking you in. You see what shipped, what it cost and what happened every month.

No. A percentage of whatever revenue gets attributed to AI invites arguments about branded search, existing customers and other campaigns. A narrow performance-linked component can be added later, once accepted enquiries are defined and reported reliably, with CRM visibility, duplicate handling, a measurement window and a dispute process agreed in writing.

The diagnostic, nearly always. It is a standalone deliverable with no obligation, and it tells you whether there is a gap worth paying to close and whether a programme could pay for itself on your economics. If you are already well represented for the questions that matter, or your enquiry handling is the real bottleneck, we will say so, and you shouldn't buy a programme.

On current evidence, no. Google has explicitly stated it ignores llms.txt, and OpenAI's, Anthropic's, and Perplexity's crawler documentation describes robots.txt-based controls without ever mentioning consuming llms.txt from third-party sites. Several of those companies publish an llms.txt for their own docs, which is not the same as reading yours. Ship one if you like — it takes an hour and is harmless — but nobody should sell it to you as a visibility lever.

Per-engine crawler access in robots.txt, and it's the one most sites get wrong. Each major engine runs several separately-blockable agents that split training, live retrieval, and search indexing. A blanket 'block AI bots' rule is a common misconfiguration that silently removes you from the citable surface — you stop being quotable in live answers, not just excluded from training data. It's free to fix and better documented than anything else in this field.

No. Google states that no structured data type is required for or specific to its generative AI features, and it fully removed FAQ rich results on 7 May 2026. Ship schema for conventional rich-result eligibility and for entity clarity — it helps machines understand who you are — but not as an AI-visibility lever. Whether markup affects citation in non-Google engines is an open question with no good public evidence either way.

That's a genuine business decision, not a technical default — but make it deliberately and per-agent. Blocking training crawlers while allowing retrieval and search agents keeps you citable in live answers while limiting training use. Blocking everything with one rule removes you from AI answers entirely. Note also that compliance is voluntary and honour-based: Cloudflare has published findings of at least one AI company circumventing robots.txt.

Less than you've been told. The idea comes from the original GEO paper, which reported gains for adding statistics, quotations, and authoritative phrasing. But C-SEO Bench re-tested seven of those exact methods and found most 'largely ineffective' and frequently negative for citation ranking — while getting your document retrieved and placed high in the model's context mattered far more. Write clearly because it serves readers and can't hurt. Don't buy it as a strategy.