Skip to content

How to conduct a GEO audit: a practical step-by-step guide

A GEO audit records how AI assistants answer your buyers' questions and works out why your brand was left out or described wrongly. Run 20 to 30 decision-based prompts with branded, unbranded and comparison groups kept apart, diagnose each missed recommendation, check technical access on key pages, then pick three fixes and retest the same prompts.

A GEO audit often arrives as a slide with a single number on it, an “AI visibility score” of 34 with a red arrow next to it. That number can’t tell you whether ChatGPT left you out because it misread your category, couldn’t verify a feature you have, or preferred a competitor for a reason you might even agree with. Each of those has a different fix.

A useful audit works from evidence instead: what the assistants said, what you can verify about why, and what to look into next. Here’s how to run one without a big research budget. The examples follow a help desk vendor selling to ecommerce teams, and every number in them is illustrative.

What a GEO audit should deliver

A generative engine optimization (GEO) audit looks at how AI assistants discover, describe, cite, and recommend your brand for the questions your customers ask. By the end you should have a recorded baseline of AI answers, a list of the factual and source gaps that matter, any technical blockers you’ve verified, and a short action plan with a way to retest it.

Keep the four kinds of observation apart, because each one proves less than it seems to:

ObservationWhat it tells youWhat it doesn’t tell you
Brand mentionYour brand appeared in a recorded answerThat it was recommended
RecommendationThe assistant picked your brand as an optionThat the buyer clicked or bought
CitationYour URL was linked as supporting evidenceThat your product was recommended
Crawler requestA bot requested a page from your siteThat the page was used in an answer

A product can be documented correctly and cited and still lose the recommendation, and a single visibility score hides exactly that case.

Step 1: define the buyer decisions you’re auditing

Start with your most valuable customer instead of a list of industry keywords. Write down their situation, the problem they’re solving, and the constraints that could change their choice. For a help desk vendor, that buyer might be a support director at an EU ecommerce company whose must-haves are a Shopify integration, multilingual ticket routing, and documented EU data handling.

Then build a small category map:

LayerQuestion to includeWhat it tests
Problem“How do we manage customer messages across email, chat, and Shopify?”Whether the assistant introduces a relevant solution category
Core category“Best customer support software for ecommerce?”Unbranded brand discovery
Adjacent category“Best shared inbox for a Shopify team?”Whether category wording changes the shortlist
Buying constraints“Which help desk suits an EU retailer needing multilingual routing and Shopify?”Recommendation fit
Product validation“Does [Brand] support Shopify order data inside tickets?”Factual accuracy only, kept apart from discovery

A prompt that sounds commercial doesn’t make the person asking it in-market. Label it as a buying scenario, a hypothesis about intent, unless it came from a real customer conversation, and keep branded prompts in their own group.

The guide to choosing AI visibility prompts covers sourcing, tagging, and prioritizing prompts in depth. For a first audit, keep the sample small enough to read.

Step 2: capture a baseline you can compare later

Start with 20 to 30 distinct prompts across problems, unbranded discovery, real buying constraints, comparisons, and branded accuracy checks. That’s a working panel and nowhere near a statistically representative sample, but it’s small enough that someone can read every answer.

Run them in the AI experiences your customers use. Record the exact prompt, the engine or interface, the date, the country and language, the full response, every brand mentioned, which ones were actually recommended, and any visible citations. Repeat the strategically important prompts on several days, too, because answers move between runs: a July 2026 survey of 45 GEO studies found substantial run-to-run variability in commercial engines, along with low overlap in the sources they cite.

A simple results sheet works:

Prompt IDSegmentBranded?Brand mentioned?Recommended?Own URL cited?Main reason
P01Broad ecommerceNoYesYesNoIntegration coverage
P02EU multilingual + ShopifyNoNoNoNoOther tools had clearer EU and Shopify evidence
P03Shopify feature checkYesYesN/AYesProduct capability described correctly

The rows are illustrative. Use “N/A” when a prompt wasn’t meant to produce vendor recommendations, so a problem-exploration answer doesn’t get scored as a missed shortlist when it never asked for one.

In Quotable, you can track prompts, review recorded responses, and inspect citations across runs. That makes collection repeatable, while the value of the audit still comes from whether you picked useful buyer questions.

Step 3: diagnose the missing recommendation

If I only had a day for an audit, I’d spend most of it here. For every commercially important prompt where your brand is missing, read the full answer rather than counting mentions, and work through three questions in order.

First, did the assistant land in the right category? If the buyer describes scattered support conversations, does the answer suggest help desk software, a CRM, a shared inbox, or only process tips? Test a second category label while holding the buyer context constant.

Second, does it know what your product does? Ask a separate, explicitly branded factual question about the requirement, then check the answer against your current documentation and confirm it with the product owner as well.

Third, what explains the shortlist? Look at the evaluation criteria the answer states and the evidence it cites. Either the assistant is prioritizing a requirement your product doesn’t meet, or the public proof for a requirement you do meet is missing, weak, or wrong.

Those answers usually sort the problem into one of three buckets, and telling them apart saves a lot of wasted work. If the fact is wrong (you support Shopify order details and the answer says you don’t), start with your product documentation and any conflicting public claims. If the fact is right and you still aren’t recommended, because the assistant favors another product on pricing or regional fit, you have a selection or proof problem that a crawl audit won’t solve. And if the category is wrong, with general CRMs recommended to a buyer who needs dedicated ticket routing, look at your category language and at the sources that define the category.

A March 2026 preprint that built a taxonomy of citation failures argues the same thing from the research side, since its fixes only worked when they matched the failure, and its authors found that “generic optimization can harm long-tail content” (Tian et al.). A missing mention on its own can’t tell you which bucket you’re in, so record each explanation as a hypothesis until the evidence supports it.

Step 4: audit the sources behind the answer

Open the URLs the assistant actually cited, and instead of counting domains, look at what each source establishes. An official integration guide can verify that a feature exists, a customer case study can show how it works in practice, and a category roundup can establish which brands buyers usually compare. Those are different evidence roles.

For your top five missed buying situations, build a register like this:

Buyer requirementEvidence in cited sourcesYour evidenceNext check
Shopify order contextProduct documentation with screenshotsLogo on integrations pageConfirm supported workflows and publish details
EU data handlingVendor security documentationGeneral privacy statementVerify exact commitments and publish accurate policy information
Multilingual routingCase study and feature documentation“Global support” marketing claimConfirm languages, routing rules, and plan limits

A cited URL proves a citation and says little about the rest of the retrieval path. Some interfaces expose query fan-outs or candidate sources, but many don’t. If you can see a real fan-out, save the exact query with its parent prompt, and if all you see is citations, don’t invent the searches that supposedly produced them.

The question worth asking at the end of this step is which requirement the sources helped the assistant verify that your own public information didn’t. Answering it turns source analysis into a documentation and content backlog.

Step 5: check technical access where it matters

Technical checks belong in a GEO audit, but one missing brand mention is no reason to start rewriting the site. Pick the pages that should support your highest-value buyer decisions (product, integration, pricing, comparison, documentation, and evidence pages), then check:

  • Does each URL return the expected status and render real content without a login?
  • Can Google index the eligible pages, and are canonical and robots directives intentional?
  • Do your server or CDN logs show verified AI crawler requests, errors, or unusual blocking on those URLs?
  • Are security rules, redirects, or client-side rendering stopping a permitted crawler from reaching essential information?

The CDN question deserves a second look if you’re on Cloudflare. Since July 1, 2025, new Cloudflare domains have blocked known AI crawlers by default (MIT Technology Review), and Cloudflare has since tightened its defaults for ad-supported pages, so a 403 in your logs may come from a setting nobody on the team chose.

OpenAI runs three separate agents (OpenAI crawler documentation): OAI-SearchBot for search, GPTBot for model training, and ChatGPT-User for certain user-triggered page requests. They’re controlled independently, so blocking GPTBot leaves you in ChatGPT search, while blocking OAI-SearchBot keeps your pages out of ChatGPT search answers. Check each one by name instead of treating every AI user agent as the same thing.

For Google, check two things. Its AI optimization guidance says generative Search features build on ordinary SEO requirements and need no special llms.txt file or AI content chunks, so foundational access comes first. Search Console also added a control in 2026 for appearing in AI Overviews and AI Mode, so confirm that nobody on the team opted the site out.

Quotable’s AI crawler analytics can examine bot activity once server or CDN logs are connected. Keep the two kinds of evidence straight, though: seeing no crawler requests doesn’t prove a bot can’t reach a page, whereas a verified 403 on an important URL is a concrete issue to fix.

Step 6: pick three fixes

By now you’ll probably have more possible work than the team can do. Rank findings by commercial impact, confidence in the evidence, and effort, and don’t dress subjective judgment up as a scientific “GEO opportunity score.”

Write one row per finding:

FindingConfidenceBusiness impactNext actionHow you’ll verify it
Shopify feature publicly unclearHigh: documentation lacks detailsHigh: buyer must-haveUpdate verified integration documentationBranded factual checks plus unchanged buyer prompts
Low category visibility for “shared inbox”Medium: observed in a few runsMediumTest more category variants before changing positioningAdditional controlled prompt sample
AI crawler blocked by CDN ruleHigh: verified 403 in logsHigh if the page is criticalFix the security rule without weakening protectionsConfirm successful permitted requests

Sometimes the best thing an audit can recommend is to publish nothing new. Fixing an inaccurate integration page, correcting a third-party listing, or repairing an access rule can do more than another generic GEO explainer.

For repetitive research and content tasks across many pages, Quotable’s Tables and agent workflows can help with execution. Keep a person responsible for checking product claims and approving anything that goes public.

Step 7: retest the original questions

Go back to the same prompts, engines, locale settings, and classification rules after the change, and keep new prompts in a separate exploration panel so harder questions don’t look like a drop in performance.

Measure the outcome the fix was supposed to change. For a factual correction, that’s how many of the verified claims you tested are now accurate. For discovery, it’s how many eligible unbranded recommendation answers name the brand. For citations, it’s how often the relevant URL is cited in the fixed panel. And for commercial value, it’s whether identified AI referrals, qualified leads, or buyer inquiries change, with the attribution limits stated plainly.

Say your brand appeared in 3 of 12 EU ecommerce recommendation answers before the fix and 7 of 12 after. That’s a move from 25% to 58% for those recorded answers. With a sample that small and AI systems that keep changing, it doesn’t prove the website update caused the increase, and it certainly doesn’t mean your market-wide visibility is now 58%. The survey mentioned in Step 2 puts this in perspective, since across the research it reviewed, “no reviewed technique shows a stable, longitudinal, cross-platform causal effect” on organic discoverability.

Keep the full answers, too. A changed recommendation that comes with a correct explanation of your Shopify support tells you more than the percentage does.

A report you can hand to your team

A first GEO audit report fits on two or three pages. Open with scope: the buyer segments, markets, engines, prompt groups, time window, repetitions, and known limitations. Then the baseline, meaning unbranded recommendation counts, branded factual accuracy, own-domain citations, and the most common reasons for selection or exclusion, always with numerators and denominators.

The middle of the report is three evidence-backed findings, each with the prompt and response, the relevant URLs or logs, a confidence rating, the recommended action, an owner, and a retest date. Close with decisions: what gets fixed now, what needs more investigation, and what you’re dropping because the evidence is weak or the opportunity is small.

If you use Google, add the Search Console generative AI performance reports that Google launched in June 2026. They show impressions in AI Overviews and AI Mode by page, country, date, and device, which is useful first-party data. They don’t include clicks or queries, though, and they tell you nothing about ChatGPT answers or brand recommendations.

GEO audit checklist

Before you call the audit complete, make sure you can answer each of these:

  • Did we select commercially relevant buyer situations and category variants?
  • Are branded, unbranded, and comparison prompts reported separately?
  • Did we save the exact answers, engines, dates, and visible citations?
  • Did we distinguish mentions, recommendations, citations, and crawler activity?
  • Did we verify important product claims against current, authoritative facts?
  • Did we inspect cited sources for the evidence they actually provide?
  • Did we confirm technical failures in logs instead of guessing from missing bot activity?
  • Does every priority finding have evidence, an owner, an action, and a retest date?

What to do next

Start with one buyer segment that matters to revenue, one carefully chosen panel, and three findings you can verify. Fix the clearest problem first, leave the benchmark alone, and see what the next round of answers tells you.

For the bigger picture, read our complete guide to generative engine optimization, and if the weak spot is the prompts themselves, start with how to choose the right prompts for AI visibility tracking.

If your audit ends with forty slides and no owner for the first fix, it isn’t finished yet. Quotable brings the prompts, responses, citations, crawler data, and follow-up workflows into one place, so the time goes into that first fix instead of the slides.

Frequently asked questions

What is a GEO audit?

A GEO audit looks at how AI assistants discover, describe, cite, and recommend your brand for the questions your customers ask. It ends with a recorded baseline of AI answers, the factual and source gaps that matter, any verified technical blockers, and a short action plan you can retest.

How many prompts do I need for a GEO audit?

Start with 20 to 30 distinct prompts across problems, unbranded discovery, real buying constraints, comparisons, and branded accuracy checks. That’s a working panel rather than a representative sample, but it’s small enough that someone can read every answer.

Why is my brand missing from AI recommendations?

Usually for one of three reasons: the assistant misread your category, it has wrong facts about your product, or it preferred another product on fit or proof. Each needs a different fix, so read the full answer and diagnose before changing anything.

Can my CDN block AI crawlers?

Yes. Since July 1, 2025, new Cloudflare domains have blocked known AI crawlers by default, so a 403 in your logs may come from a setting nobody on the team chose. Check each crawler by name, since OpenAI’s search and training bots are controlled separately.

How do I know if my GEO fixes worked?

Retest the same prompts, engines, and locale settings, and measure the outcome each fix was meant to change, such as factual accuracy or unbranded recommendations. With small samples, treat changes as observations: moving from 3 of 12 to 7 of 12 answers doesn’t prove the fix caused it.

What should a GEO audit report include?

Two or three pages: the scope (segments, markets, engines, prompt groups, time window), a baseline with numerators and denominators, three evidence-backed findings each with an owner and retest date, and decisions on what to fix now, investigate further, or drop.