An SEO audit checks four areas to see whether your website can be found, understood, and delivered quickly: technical setup (reachability, indexability, load time), on-page work (titles, descriptions, headings, links), structured data, and, for a local business, local relevance. That is the base audit, and on a small, technically simple site free tools carry all of it: Search Console, Bing Webmaster Tools, Lighthouse, and Screaming Frog free up to 500 URLs. Without a measurement tool attached, the AI does not collect those values itself, and that is exactly where most guides get sloppy. It reads the raw output of your tools, finds patterns across hundreds of pages, and turns that into a prioritised action plan. Tools measure, AI interprets. Ask the AI instead for "my site's SEO score" and you get a number that came from nowhere. A pass through all seven steps costs about half a day on a site under 500 URLs, and it ends with the follow-up measurement, not with the report.
I run my business as a solo freelancer with a workforce of AI employees built on Claude, and one of them does nothing else: measure, assess, measure again. How the concept is built is on AI employees, the roles one by one in My AI Workforce. This piece is about the process, including the parts where AI contributes nothing.
What an SEO audit actually checks
An audit is not an opinion, it is a list of findings with a source. Four areas cover what goes wrong on small sites.
These four areas are the base audit for a small, technically simple website. Not in it: backlinks and off-page signals, multilingual sites, JavaScript rendering, log files, relaunch migrations, complex information architecture. Whether one of those is the decisive point for you does not depend on your page count. A clean audit therefore fixes the scope up front and writes it into the report, instead of stamping "fully checked" underneath at the end.
FAQPage is a special case: the markup can be valid, but according to Google's own documentation, FAQ rich results are now shown only for well-known, authoritative government and health websites. For an ordinary business it is clean markup without a special effect, and that is how it belongs in the report.
What an audit does not deliver is a promise about where you will rank. Google in its own starter guide: "There are no secrets here that'll automatically rank your site first in Google (sorry!)", and further: some changes take effect in hours, others take months. Anyone promising you a position contradicts the company that builds the search engine.
The free tools that are enough for an audit
Paid SEO suites are rarely the bottleneck for a one-person business. The bottleneck is that nobody measures. These six tools cover the process; how they stack up against crawler licences and third-party index suites is in the tool comparison.
Most audits fall apart at the difference between lab and field data. Lighthouse measures under controlled conditions; the Chrome UX Report is, in Google's words, "a dataset that reflects how real-world Chrome users experience popular destinations on the web", and it does not contain every page, because a page needs enough visitors to qualify. On a new site you will therefore often have lab data only, and that belongs in the report as such, rather than passing lab numbers off as user experience.
An SEO audit in seven steps
1. Sort out access and define the goal
Without Search Console you are measuring in the dark. Set it up if it is missing, and write down in one sentence what the audit should answer: "Why does the services page get no clicks?" is a goal, "do some SEO" is not. If someone else audits for you, they get read access and nothing more.
2. Crawl the site
Screaming Frog pulls up to 500 URLs for free. Then run linkinator over the same domain for broken links and chains. One thing matters: keep the rate moderate. On cheap shared hosting, a crawl that is too fast takes the site down for half an hour. One request every three to five seconds is my house rule for sites that are not mine, not a guarantee: a crawl delay in robots.txt, the terms of use, and whatever the server can take all come first.
3. Check indexability
Fetch robots.txt and sitemap.xml and hold them against each other: does the sitemap list a URL that returns 404 or carries noindex? Does robots.txt accidentally block a whole directory? The Search Console indexing report shows the same question from Google's side, with reasons for excluded pages.
4. Measure load time, field before lab
The three Core Web Vitals and their thresholds: Largest Contentful Paint up to 2.5 seconds, Interaction to Next Paint up to 200 milliseconds, Cumulative Layout Shift up to 0.1. Assessment happens at the 75th percentile, split by mobile and desktop. Measure your three to five most important URLs, not just the homepage, field and lab data in separate columns.
5. Work through on-page and structured data
Title and description lengths, duplicates, and missing h1 tags fall out of the crawl on their own. Then extract the JSON-LD of your main page types and check it against Schema.org. This is where AI saves real time for the first time: scanning 500 lines of crawl output for patterns is its strength.
6. Read the Search Console data against it
Now real user data enters. What matters are queries at position five to fifteen with many impressions and few clicks, and queries where several of your own pages show up at once. The second case is a reason to diagnose, not yet a diagnosis: behind it sit overlapping search intent, texts that are too similar, a wrong canonical, the internal linking, or old URLs still in play. The Search Console rows alone do not settle it. Take the affected pages and check content, intent, canonicals and internal links together, then the cause is clear.
7. Action plan and follow-up measurement
Every action gets a finding, an effort estimate, and an expected effect as a hypothesis, not as a promise. No action without a finding, otherwise you are copying the textbook. And a date for the follow-up measurement, because an audit without a second measurement is an opinion with tables.
What gets measured and what the AI interprets
A language model has no browser, no crawler, and no access to your Search Console unless you give it one. Reading, comparing, prioritising it can do. The table splits the two per step.
The column that counts is the first one. In no step does the measurement come into being inside the language model, it comes into being in the tool. Whether you start it or an agent runs it changes nothing about that, one of them just has to run. The order is always: run the tool, save the output, hand it to the AI, get the assessment back. Written down as a skill, you never reinvent the prompt: Claude Skills: How to Create and Use Them.
Where AI produces nonsense in an SEO audit
Four mistakes come up again and again.
Invented measurements. An AI without tool access answers every question about numbers with a number: ask it for your Core Web Vitals and you get an invented one. The countermeasure is banal and it works: every figure in the report points at a saved file. No file, no figure.
Textbook actions. Ask an AI for SEO tips in general and you get the same twenty points everyone else gets. Ask it to assess your crawl output and you get five that fit your site. The difference is the data you hand over.
Ranking promises. Google itself: "While there's no guarantee that any particular site will be added to Google's index, sites that follow the Search Essentials are more likely to show up in Google's search results." An AI that writes "page one in three months" into the plan made that up.
Trends without evidence. The best known example is llms.txt, meant to tell AI systems which pages count. Google's John Mueller on 15 June 2026: "It's basically you're telling these systems, like, I have the best website ever. And here are all of the pages that everyone must go to." His only concession: "If someone is already on your website, maybe some kind of automated system is helpful." An Ahrefs analysis of 137,210 domains from June 2026 fits that picture: of roughly 38,000 domains with a valid llms.txt, 97 per cent received no request for the file at all in May 2026. Neither benefit nor harm is evidenced; the only defensible statement is that there is no demonstrated priority lever. It does not belong in an action list where every line has a finding and a reasoned effect.
The employee who runs this process on a schedule
The difference between an audit and ongoing work is not the depth of the report, it is the repetition: whether an action worked is something only the second measurement tells you, and that is the part most often skipped. This is why the process is a permanent role here, not a project. Sebastian, my AI employee for SEO and GEO, takes a URL and a goal, runs the seven steps with the same tools, files every raw output in a dated job folder, and delivers an assessment in which each finding names its source. He does not change live sites, and in accounts that are not mine he has read access only. He measures, I decide.
This role also measures AI search. The "AI Performance" report in Bing Webmaster Tools has shown since February 2026 how often Microsoft Copilot cites individual pages. On my own site that produced the finding that justifies the way of working: the page with the most Google clicks and the one with the most Copilot citations were not the same. Ranking and citation frequency are separate measurements; measure only one and you see half the picture.
If you would rather build such a role yourself, the path is in Creating an AI Employee: How to Start. For a business still working out where AI pays off at all, Bringing AI to a Small Business is the entry point.
Frequently asked questions
Can an AI analyse my website for SEO?
Assess yes, collect only with a tool attached: without a crawler, without Lighthouse, without access to your Search Console, a language model has no data source. The tools supply the raw data, the AI reads it and writes the assessment and action plan. Ask directly for numbers and you get invented ones.
What does an SEO audit cost with free tools?
Nothing but your time. Search Console, Bing Webmaster Tools, Lighthouse, and PageSpeed Insights are free, Screaming Frog crawls up to 500 URLs for free, linkinator is open source. For a small, technically simple site under 500 pages that carries the whole base audit. Backlink analysis, JavaScript rendering, log files, or a relaunch migration need more, and that is the point at which the question arises whether you do this yourself or commission an agency. The Screaming Frog licence, which unlocks JavaScript rendering, saved crawls and crawl comparison among other things, costs 245 euros per year (as of September 2026).
How often should I run an SEO audit?
The whole process once a year or after a relaunch, the follow-up measurement for individual actions four to eight weeks after they go live. Google itself writes that some changes take effect in hours and others take months. In between, check Search Console monthly: its data reaches back 16 months and then disappears.
Do I need Screaming Frog, or is Lighthouse enough?
Both, they measure different things. Lighthouse inspects a single page in depth across performance, accessibility, best practices and SEO. Its additional Agentic Browsing category is experimental, requires Chrome 150 or newer and needs the Origin Trial for WebMCP checks. Screaming Frog goes wide and finds what is only visible by comparing many pages: duplicate titles, redirect chains, orphaned pages. Free covers 500 URLs and the basic data this process needs; JavaScript rendering, saved crawls and crawl comparison depend on the licence.
Does llms.txt help visibility in AI search?
There is no evidence for it so far. Google's John Mueller classified the file in June 2026 as self-reported information that does not help an AI system tell websites apart. An Ahrefs analysis found that 97 per cent of the llms.txt files it examined received no request at all in May 2026. Neither benefit nor harm is evidenced; the only defensible statement is that there is no demonstrated priority lever. Anyone creating it should do so of their own accord, not because an offer bills it as a measure. What is on the list instead, crawler access, readable HTML and one page per question, is a guide of its own.
How to take the next step
Start with one measurement rather than the whole audit: set up Search Console if it is missing, and run Lighthouse over your most important page. Twenty minutes, and for the first time you have numbers instead of guesses. Then the crawl, and only then the action plan, in the order that counts for a small business.
If you do not want to run this process yourself every quarter, the role matters more than the tool: AI employees and My AI Workforce. The templates, the checklist for these seven steps, and the people who are working through the same questions right now are in my community.
