Jean-Romain
People now ask AI the questions that used to go to Google. ChatGPT alone is reportedly above 900 million weekly active users, so "recommend me the best X" is becoming a default buying behavior. Google is also injecting AI answers straight into the SERP: one large study found AI Overviews stabilized around 16% of queries by late 2025. You can rank on page one and still lose the demand if AI does not mention or cite you.
An AI SEO audit is how you check whether AI assistants can reach, read, recognize, and answer from your site, then fix what blocks them. It is not a traditional SEO audit with a new coat of paint. A classic audit asks "can Google rank this page". An AI SEO audit asks a harder question: can ChatGPT, Perplexity, and Google AI Overviews crawl your pages, parse them without choking on JavaScript, understand who you are, and quote you as the answer. I run this exact process on client sites and on awilix.ai itself. Here is the whole thing, step by step.
What an AI SEO audit checks, and how it differs from a traditional SEO audit
Answer first: an AI SEO audit scores four things an AI system has to do before it can recommend you. Miss one and you stay invisible in AI answers no matter how well you rank in blue links.
- Reach: can AI crawlers fetch your pages at all (robots.txt, sitemap, llms.txt, no accidental block on GPTBot, ClaudeBot, PerplexityBot, Google-Extended)
- Read: can they parse the content, or is it trapped behind JavaScript that only renders in a browser
- Recognize: do they understand your entity (schema, consistent positioning, clear "who you are" signals)
- Answer: is your content easy to lift and quote in one clean block, backed by proof
This is the GEO layer (getting cited in ChatGPT and Perplexity answers) sitting on top of classic SEO, not a replacement for it. If you want the full checklist of the classic layer too, see what a professional SEO audit includes.
| Dimension | Traditional SEO audit | AI SEO audit |
|---|---|---|
| Core question | Can this page rank on Google | Can AI reach, read, and cite this page |
| Main signals | Keywords, backlinks, Core Web Vitals, titles | Crawler access, rendering, schema, answer blocks, entity clarity |
| Where results show | Blue links | ChatGPT, Perplexity, Gemini, AI Overviews |
| How you measure | Rankings and clicks in Search Console | Mention, citation, and framing across LLM answers |
| Biggest risk if ignored | Lower rankings | Invisible in AI answers while still ranking |
The rest of this guide is the exact run order: pick the keywords that matter, test them on real models, diagnose the gap, fix the technical layer, then ship. If you would rather I run it for you, the free SEO and GEO audit lands in 24 hours.
Step 1: Select your money keywords
An AI SEO audit is only useful if it is built on keywords tied to revenue. Skip the "let's audit everything" approach, you will waste time and get noisy results.
Your objective: 50 to 200 keywords that represent what makes money now and what should make money next.
Goal: Define the scope fast
You want keywords that answer:
- What brings leads or sales today
- What buyers search right before choosing
- What AI should mention when people ask for recommendations in your niche
Where to pull keywords: The fastest sources
- Google Search Console: queries from converting pages, last 3 to 6 months
- GA4: landing pages with conversions, map to their main queries
- Paid search: best converting ad groups, pure intent
- Sales calls: exact phrases customers use
- Competitors: 5 to 10 brands customers compare you with
Keep or drop: Simple filter
| Keyword type | Keep when | Drop when | Example |
|---|---|---|---|
| Money intent | Can convert | Pure curiosity | "AI SEO audit service" |
| Problem intent | Clear pain | Too generic | "not cited in ChatGPT" |
| Category intent | Purchase decision | Too broad | "AI SEO tools" |
| Brand intent | Comparison or reviews | Noise | "Brand drama" queries |
Bucket keywords: Avoid overlap
Create buckets so prompts don't repeat the same test.
- Recommendation: best, top, agency, tool
- Comparison: vs, alternative
- Use case: for SaaS, for ecommerce, for Shopify
- Fix: how to, why not, troubleshoot
Step 2: Turn keywords into prompts that match real AI conversations
Keywords are SEO language. Prompts are buyer language. Your job is to translate each keyword into the question someone would actually ask ChatGPT.
Also: LLM outputs can vary across runs, even with similar inputs, so you need consistent prompt formats to compare results cleanly.
Goal: Build a prompt library that produces stable comparisons
Each prompt should make it easy to log:
- Mention: are you named
- Citation: are you linked or sourced
- Framing: what AI says about you vs competitors
Prompt templates: Use only these 5
| Prompt type | Best for | Template |
|---|---|---|
| Recommendation | Who gets recommended | What are the best [category] for [use case] |
| Comparison | Who wins head to head | Compare [brand A] vs [brand B] for [use case] |
| Alternative | Who replaces a leader | What are the best alternatives to [brand] for [use case] |
| Use case | Niche qualifiers | Best [category] for [industry] with [constraint] |
| Fix | Visibility issues | Why is [brand] not showing up for [intent] |
Prompt rules: Make outputs easy to log
Clear and specific prompts improve output quality and consistency.
- Ask for a ranked list
- Force a format: 5 items, 1 line each
- Add context: country, budget, industry
- Keep 2 to 4 prompts per keyword max
Step 3: Run the prompts on real LLMs and log the output
This is where your AI SEO audit becomes measurable. No vibes, no "I think we're visible", just a repeatable test you can rerun next month and compare.
LLM answers change depending on the model, the day, and even the same prompt run twice. OpenAI documents this clearly: chat outputs are non deterministic by default.
Goal: Get a baseline you can compare later
You want to capture 3 signals for every prompt:
- Mention: is your brand named
- Citation: is your brand linked or used as a source
- Framing: how the model describes you vs competitors
Model selection: Use what people actually use
Start with the assistants that dominate usage. StatCounter's Jan 2026 snapshot shows ChatGPT leads by a lot, followed by Perplexity and Gemini, then Copilot and Claude.
Recommended test set:
| LLM | Why it matters in an AI SEO audit |
|---|---|
| ChatGPT | Highest usage share, sets the "default" narrative |
| Google Gemini | Connected to Google ecosystem, strong for discovery prompts |
| Perplexity | Citation heavy outputs, great for source footprint checks |
| Microsoft Copilot | Common in enterprise workflows |
| Claude or Grok | Claude is common in B2B, Grok is rising fast in the US |
Pick 5 and stick to them for the whole audit. Consistency beats perfection. For the ChatGPT side specifically, I go deeper in how to do SEO for ChatGPT.
Test rules: Keep conditions stable
Small changes create fake insights. Use a strict setup:
- New chat for each prompt
- Same language and same country context
- Same browsing mode setting across runs
- Run each prompt 2 times if results look unstable
- Save proof: screenshot or copy the raw output into your log
Awilix can run your AI SEO Audit end to end, then turn the findings into a prioritized roadmap that increases rankings, LLM mentions, and citations. Book a call with us now
Step 4: Diagnose the gap: mentions vs citations vs narrative
Raw LLM answers are messy. This step turns them into clean diagnosis.
Also a quick reality check: AI visibility is volatile. One 2026 tracking study reports only 30% of brands stayed visible from one answer to the next, and only 20% stayed present across five consecutive runs. That is why you diagnose patterns, not single screenshots. If you want a repeatable scoring method, I detail one in how to measure brand visibility in AI search.
Goal: Convert outputs into 3 clear buckets
For each prompt cluster, you want one of these outcomes:
- Not mentioned
- Mentioned not cited
- Cited but weak narrative
Once you have the bucket, the fixes become obvious.
Deliverable: Gap map table
This is the table that creates your roadmap in 10 minutes.
| Prompt cluster | Outcome | What it usually means | What to check next |
|---|---|---|---|
| Recommendation | Not mentioned | You are not considered a default option | Source footprint, entity clarity, coverage |
| Recommendation | Mentioned not cited | You exist, but AI does not trust your site as a source | Proof signals, extractability, authoritative pages |
| Comparison | Cited but weak narrative | You are used, but framed poorly | On page positioning, claim proof, specificity |
Bucket 1: Not mentioned: You are not in the model's shortlist
This is rarely a "page tweak" problem.
Check these levers first:
- Source footprint: do trusted sites mention you at all
- Entity clarity: is your category and positioning consistent across your site
- Coverage: do you have pages that match the prompt intent cleanly
Bucket 2: Mentioned not cited: You are known but not trusted as a source
This is the most common pain.
Typical causes:
- Your pages are hard to quote: no clear answer blocks, weak structure
- Your claims have no proof: no data, no sources, no specifics
- The model relies on other domains that already summarize your niche
Bucket 3: Cited but weak narrative: You are visible but framed wrong
This is the sneaky one. You "win" citations but lose preference.
Fixes usually live here:
- Make your positioning explicit in plain sentences
- Add concrete differentiators: constraints, use cases, pricing logic, outcomes
- Remove vague marketing language that AI rewrites into bland generic text
Mini experiment: The quickest way to spot your real problem
Pick 10 prompts where competitors show up.
If you are:
- Missing everywhere: you need presence across trusted sources
- Mentioned but never linked: you need citation ready pages and proof
- Linked but described poorly: you need tighter positioning and on page clarity
That is the whole point of an AI SEO audit: fewer guesses, more diagnosis, faster roadmap.
Step 5: Audit the technical layer so AI can reach and read you
This is the step most brands skip, and it is the one that quietly makes every other step pointless. If an AI crawler cannot fetch your page, or the content only appears after JavaScript runs, nothing you write gets into the model. Google is explicit that there are no special tricks for AI Overviews, the fundamentals decide eligibility, and the same fundamentals decide whether ChatGPT and Perplexity can use you.
Check 1: AI crawler access
Open your robots.txt first and confirm you are not blocking the bots that feed AI answers. A single stray Disallow can hide your whole site from a model.
| Crawler | Who it feeds | Default call |
|---|---|---|
| GPTBot, OAI-SearchBot | ChatGPT and ChatGPT search | Allow unless you have a specific reason not to |
| ClaudeBot | Claude | Allow |
| PerplexityBot | Perplexity | Allow |
| Google-Extended | Gemini and AI Overviews | Allow to stay eligible |
| Googlebot | Classic index and AI Overviews | Always allow |
Check 2: Rendering and machine readability
View source on your money pages, not the rendered DOM. If the main content only shows up after JavaScript executes, many AI crawlers never see it. Server side rendering or static output beats client side rendering for this.
- Is the core copy present in the raw HTML, not injected client side
- Is the heading hierarchy clean (one H1, logical H2s) so a model can map the structure
- Is the content to script ratio healthy, or is the page mostly framework code
Check 3: Structured data and entity signals
Schema tells a model what each thing on the page is. Validate structured data and keep it aligned with the visible content.
- Organization schema with consistent name, logo, and sameAs links so the model recognizes your entity
- Product or Service schema with pricing where it applies
- FAQ schema on pages that answer buyer questions
Entity clarity carries beyond schema. If you want the deeper version, see entity SEO, and for the classic crawl and index fundamentals use the technical SEO audit checklist.
Check 4: The map files
- XML sitemap: current, no 404s, submitted in Search Console
- Canonicals: every key page points to itself, no accidental loops
- llms.txt: a curated map of your best pages, useful for documentation heavy sites and large blogs (not an official standard yet, adoption varies)
Deliverable: an "eligibility blockers" list ranked by severity: blocker, affected URLs, fix, owner. Fix indexing and crawler blocks first. Everything else is wasted effort until a model can actually reach the page.
Step 6: Build the roadmap: 4 levers that move AI visibility
Now you have the gap map and the technical blockers. This step turns them into actions you can ship. Think of it like an AI SEO audit scoreboard: you fix what blocks visibility first, then what improves LLM mentions and citations at scale.
Lever A: Source footprint: where LLMs already pull answers from
If you want citations, you need to win the places AI already trusts. Perplexity is the easiest example because it shows citations by design.
This is also the lever with the clearest proof. We ran a PR campaign engineered for LLM visibility rather than raw backlinks, choosing placements for their odds of being ingested by models. It produced 22M potential reach, 565 placements, and a 25% detection rate on LLMs. The full story is in the PR campaign for AI visibility case study.
What to do
- List the domains that show up most often in citations for your prompt clusters
- Tag them by type: media, directories, forums, review sites, industry blogs
- Pick the 10 that appear repeatedly, those are your target surfaces
Quick wins
- Create 1 page those sources can link to cleanly: "What we do" plus proof plus FAQs
- Pitch 3 to 5 niche publications that already rank for your category
Lever B: On page extractability: can your site be quoted in one clean block
This is the "make it easy to cite" lever. Not prettier content, just more usable content.
What to check
- Does the page answer the question fast, then expand
- Can a model lift a definition, steps, or comparison without rewriting everything
- Are lists and tables doing real work, not decoration
Quick wins
- Add 3 short answer blocks: definition, best practices, checklist
- Add 1 comparison table if the query is recommendation based
Lever C: Technical eligibility: close the blockers from Step 5
You already found these in the technical audit above. Here they become roadmap items: unblock the AI crawlers, fix client side rendering on money pages, ship the missing schema. This lever is low glamour and high leverage, so it goes to the top of the backlog.
Lever D: Entity and authority: do you deserve to be recommended
This is where you stop being "a random option" and become "a default pick".
What to do
- Make your positioning unmissable in plain sentences
- Close content gaps around your highest value prompt clusters
- Add internal links that connect your best proof pages to your money pages
Quick wins
- Add a 5 sentence "entity block" on key pages: who you help, what you do, what you are best at
- Publish 2 supporting pages that answer buyer sub questions, then link them into the main page
Step 7: Execute fixes: 14 day sprint plan
A roadmap is useless until it ships. The easiest way to execute an AI SEO audit is to run it like a sprint: short, focused, measurable. Scrum defines a Sprint as one month or less, most teams operate faster than that because feedback loops matter.
Goal: Ship changes that improve mention and citation rate
Your sprint needs one rule: every task must change an output you can retest.
Good sprint outputs:
- Fewer eligibility blockers
- More citation ready pages
- Better source coverage on target domains
- Better narrative when the model compares you to competitors
Sprint setup: What goes in the backlog
Keep the backlog small. If it cannot be done in 14 days, it is not a sprint item.
| Backlog item | Lever | Impact | Effort | Definition of done |
|---|---|---|---|---|
| Fix indexation blocker on key page | Eligibility | High | Low | Page indexed and canonicalized correctly |
| Add answer blocks to 5 money pages | Extractability | High | Medium | Clear TLDR plus steps plus table live |
| Publish 2 support pages for 1 cluster | Authority | Medium | Medium | Pages live plus internal links added |
| Get featured on 3 recurring cited domains | Source footprint | High | High | Live mention plus link or profile live |
Prioritization: Use impact vs effort
Do not debate priorities for hours. Use a simple framework, value vs effort is one of the most common because it forces tradeoffs fast.
| Quadrant | What you do |
|---|---|
| High impact: Low effort | Do first, same week |
| High impact: High effort | Plan, assign owner, ship partial if possible |
| Low impact: Low effort | Do only if time left |
| Low impact: High effort | Kill it or park it |
Want the fastest path to more LLM mentions and citations: book a call and Awilix will run your AI SEO Audit, then ship the fixes with you in a 14 day sprint. The first audit is free and lands within 24 hours.
FAQ on AI SEO Audits
How do I check if AI crawlers can access my site
Open yourdomain.com/robots.txt and look for any Disallow that hits GPTBot, OAI-SearchBot, ClaudeBot, PerplexityBot, or Google-Extended. Then view source on a key page and confirm the main content is in the raw HTML, not injected by JavaScript. Those two checks catch most "we are invisible in AI" problems before you touch content.
How many prompts do I need for a reliable AI SEO Audit
For most brands, 60 to 120 prompts is enough to spot patterns. If you cannot explain your results in 5 minutes, you tested too much, not too smart.
Should I test with browsing enabled
Use both.
- Browsing off: tests model memory and training signals
- Browsing on: tests real time retrieval and citations
Keep them separated in your log or you will mix two different realities.
How do I avoid fake citations from LLMs
Treat citations like leads, verify them.
- Open the URL
- Check the page actually mentions you
- Confirm the claim appears on the page
- If it is wrong, mark it as hallucinated and ignore it
What is llms.txt and when it helps
llms.txt is a proposed file to help LLMs pull the right pages at inference time, basically a curated map of your best content. It is not an official standard and adoption varies.
Best use cases:
- Documentation heavy sites
- API products
- Large blogs where models need a shortcut to "the right pages"
Why do some models mention me but never cite my site
Because the model trusts a third party summary more than your pages.
Quick fix that often works: publish a "facts and positioning" page that is easy to quote:
- 5 plain sentences on what you do and who it is for
- proof points with sources
- a short FAQ with crisp answers
