Frequently asked questions about GEO audits

GEO stands for Generative Engine Optimisation: the work of getting your business retrieved, described accurately, and cited inside the answers AI systems generate. A classic SEO audit asks whether a page can rank. A GEO audit asks whether a page can be retrieved, extracted, attributed and cited. Those are related questions with different answers, and the gap between them is wider than most people expect.

A GEO audit is a structured diagnosis of why AI answer engines are not recommending you. An GEO audit covers 52 points across seven sections: AI crawler and retrieval access, the classic SEO foundations that gate it, passage and chunk structure, entity and authority, structured data, prompt visibility testing across seven AI surfaces, and measurement and monitoring. The output is an action plan tiered by impact and effort, plus the full log of every prompt we ran and the list of sources the engines actually cite in your category.

An GEO audit costs $990. There is one price and one tier. See the full pricing page for what's included.

Five to eight business days from intake to delivery, with 10 to 14 hours of manual audit work inside that window before the report is written. The prompt set has to be built, agreed, run and repeated before the diagnostic work starts. We don't promise 24-hour audits, because a single pass across a handful of prompts tells you almost nothing.

Each GEO audit is delivered as a single PDF containing an overview, all 52 checks scored and explained, the issues ranked by severity, an action plan tiered from fix-now to de-prioritise, implementation notes for the highest-impact findings, the complete prompt log, and the cited source target list. An optional 1h walkthrough call is included. If you want the working files behind the report, ask and we will send them. See what's delivered for the full list.

No, but it is taking a share of the traffic SEO used to capture. AI Overviews now appear on roughly 25% of Google queries, and when one appears, click-through on position one falls from 1.41% to 0.64%. The ranking still matters; it just gets read less often. Most businesses need both, and several fixes serve both at once.

Not necessarily. Only a minority of the URLs cited by AI engines also sit in Google's top 10 for the same query. Engines lean heavily on sources classic SEO barely touches: community threads, review platforms, industry publications and independent roundups. A site can rank first and be invisible to ChatGPT, and a site can be cited by Perplexity while ranking on page three. Plenty of category leaders are absent from the answers about their own category.

Monitoring tools that track brand mentions across engines are useful and getting better, and if you want ongoing tracking you should buy one. What they do not do is tell you why you are absent. A tool reports that a competitor was named in 8 of 10 runs. It does not tell you that a CDN rule is returning 403 to PerplexityBot, that your priority pages bury their answer four paragraphs down, or that both of that competitor's citations came from one roundup article you could be added to in a week.

It depends on whether your buyers use AI to shortlist. In most B2B and considered-purchase categories they increasingly do, and the field is uncontested enough that the fixes are cheap relative to the position they win. In categories where buying is impulsive, local, or purely price-driven, the case is weaker. We'll tell you if we think you're in the second group.

You decide. You can take the report and execute everything in-house, or scope an implementation project with us for the parts you would rather hand over. The audit is written to stand on its own; nothing about it requires further engagement.

ChatGPT as the primary platform, since it accounts for the large majority of AI referral traffic, then Perplexity, Gemini, Claude and Copilot. Google AI Overviews and Google AI Mode are tested separately, because they are different surfaces with different citation patterns. Every prompt runs logged out, with personalisation and memory off, two to three times, because a session that already knows your brand produces results no prospect would see.

Yes, and the differences are large enough to change your priorities. Perplexity is citation-heavy and links generously. Copilot grounds on the Bing index, which is why Bing indexation gets its own check even on sites where Google indexation is perfect. Google AI Overviews and AI Mode diverge from each other regularly: a site cited in the Overview is often absent once the user asks a follow-up in AI Mode. The audit scores each surface separately rather than averaging them into one number.

Because it reads as unfiltered human experience, it is dense with the exact phrasing people use when they describe problems, and it is licensed and accessible to several engines. Reddit is currently the single most-cited domain across ChatGPT, AI Mode, Gemini, Perplexity and AI Overviews. That is uncomfortable if your marketing plan assumed the answer would come from your own site, and it is a big part of why the audit looks off-domain.

Not in the way you can pay for a search ad, and anyone selling you a guaranteed placement is selling something else. Ad formats inside AI surfaces are arriving, but the organic answer itself is earned through the same things the audit measures: accessibility, structure, entity clarity, and third-party citations.

It's common, and it's fixable more often than people assume. Incorrect pricing, discontinued features, or a competitor's capabilities attributed to you usually trace back to a stale page, an out-of-date third-party profile, or an old press mention that still ranks well in the engines' source set. The audit identifies the source of the error, which is the part that makes it correctable.

Yes. Answers vary by account history, region, model version, and simple randomness. That is why every prompt runs two to three times, logged out, with personalisation and memory disabled, and the result is recorded as a frequency rather than a verdict. A single screenshot of a favourable answer proves nothing, in either direction.

It depends which class you mean, and that distinction is the single most common thing sites get wrong. Retrieval bots (OAI-SearchBot, Claude-SearchBot, PerplexityBot, Bingbot, Amazonbot) build the index an assistant searches at answer time; blocking them removes you from the candidate set for citations entirely. User-triggered fetchers (ChatGPT-User, Claude-User, Perplexity-User) fire because a live person asked an assistant to read your page; blocking them means a prospect gets told the assistant cannot access your site. Training crawlers (GPTBot, ClaudeBot, Google-Extended) feed future models and are a consent and licensing question with little bearing on citation today. If AI visibility is a goal, keep the first two open and decide the third deliberately.

We check it, and we score it honestly, which usually means telling clients not to worry about it. As of 2026 no major AI provider has publicly committed to consuming llms.txt for retrieval or ranking, log analyses across very large samples find it is almost never fetched by answer bots, and Google has stated on the record that it does not support it. The one real use case is developer documentation pulled on demand by coding assistants, and if you have docs it is worth shipping. What we do flag as a problem is the implementation that generates indexable Markdown mirrors of your pages, because that creates duplicate content at scale for a speculative upside.

More than it affects Google. Googlebot renders JavaScript; several AI crawlers do not, or do it inconsistently. A client-rendered page that Google indexes fine can arrive at an answer engine as an empty shell. Server-side rendering, or at minimum static HTML for the pages you want cited, is one of the highest-leverage fixes in the whole audit.

Indirectly but meaningfully. Schema does not force a citation; it removes the ambiguity that prevents one. Organization markup with a populated sameAs array is what lets an engine connect your site to your external profiles and treat them as one entity. Person schema with sameAs is the verification pathway for author expertise, and it is consistently one of the highest-return fixes available because it takes an hour or two. Worth knowing: Google retired the FAQ rich result in 2026 and removed FAQ support from its testing tool, so FAQ markup is now purely an AI extraction signal and needs validating through Schema.org instead.

Yes, and it is one of the more frustrating findings because robots.txt looks perfectly correct while the requests never reach your server. Bot-management presets on Cloudflare, Akamai and similar services categorise AI agents as unwanted automation by default, and some now gate them behind pay-per-crawl. Checking robots.txt alone will not surface it. You have to look at what the edge is actually returning, and at server logs to see which agents are getting 403s.

If you want engines to answer usage and integration questions with your content, yes. Docs and help centres are among the most-quoted material in technical categories, and gating them or marking them noindex hands those answers to whoever left theirs open. There are legitimate reasons to keep some of it private; the audit flags the cost so the choice is informed.

It happens. There is a property-level toggle in Search Console that removes your content from AI Overviews, AI Mode and generative Discover features, and it is separate from Google-Extended and separate from ranking. Snippet directives do the same job more quietly: nosnippet, a low max-snippet value, or data-nosnippet attributes wrapping meaningful content all remove snippet eligibility, and snippet eligibility is a prerequisite for appearing in Google's AI features at all. Reversing any of this has no effect on standard Search ranking, so the decision can be made purely on AI visibility grounds.

More than its direct traffic suggests. Microsoft Copilot grounds its answers on the Bing index, and several other assistants have used Bing as a retrieval backend. Bing indexation is routinely neglected on sites where Google indexation is healthy, which means a whole retrieval surface is missing content that is otherwise perfectly available. Setting up Bing Webmaster Tools and submitting the sitemap is a short task with a disproportionate return.

As a frequency across a fixed prompt set. Ten to fifteen prompts, each run two to three times on each surface, logged out. We record how often the engines know you and describe you accurately, how often you surface unprompted when the brand is not named, how often you are cited per surface, and how often each competitor appears instead. That produces a baseline number per prompt and per engine, which is the only way to tell later whether anything actually improved.

No, and neither can anyone else. There are no positions to buy or rank for, answers are regenerated on every request, and the models change without notice. What the audit delivers is a diagnosis, a ranked remediation plan, and a measurement baseline. Where a category's answers are built almost entirely from third-party sources you have no presence on, the honest answer is that closing the gap takes months of off-site work, and the report says that rather than implying a configuration change will do it.

It depends on the cause, and the spread is wide. Access fixes such as unblocking a retrieval crawler can show up quickly, because the engine simply needs to fetch pages it previously could not. Structural content changes tend to appear in citation testing within roughly four to eight weeks, which is the sort of cycle retrieval indexes refresh on. Entity and off-site work takes considerably longer, and where the finding is absence from the sources an entire category is built on, plan in months rather than weeks.

Three layers, and the audit sets up all of them. Search Console's Generative AI report, launched June 2026, shows impressions inside Google's AI surfaces, with the page-level breakdown being the useful part; it gives no clicks, queries or position, and blends three surfaces into one. GA4 gained a native AI Assistant channel in May 2026, though it does not recognise Perplexity, so a custom channel group is needed alongside it. Server logs show the crawler side. And the prompt set is yours to re-run, which remains the most complete signal available.

The early evidence says the visits are fewer but better qualified, since the visitor has already read a synthesised comparison before clicking. Treat the published multiples with caution; sample sizes are small and the channel moves quickly. Be aware of the measurement gap too: a large share of AI sessions arrive with no referrer at all, mostly from mobile apps, and land in Direct where no tool can separate them. What is not in doubt is the zero-click side. Buyers are forming shortlists inside these tools whether or not they ever click through.

More often than you would re-audit for SEO. A full re-audit quarterly is the sensible cadence, because retrieval systems change often enough that these findings carry roughly a three to six month shelf life. Between audits, re-run the fixed prompt set monthly; it takes under an hour and it is the only way to distinguish work that produced results from work that did not. Trigger an unscheduled re-audit after any CMS migration, template revision or theme change, since those silently break schema and entity signals.