At a glance: AI tools do not read your website, they reconstruct you from every source they can find
I ran wisepreneurs.com.au through a free check that asks four AI platforms the same question about my business.
Perplexity returned fourteen checkable facts. Claude returned none, and scored me below zero. Same website, same day, four different accounts of what I do.
The gap is not a website problem, it is a corroboration problem, and it is the same problem sitting on most experienced professionals' practices right now.
This article covers what an AI agent actually reads, the four questions it works through before it recommends anyone, and the unglamorous consistency work that decides the answer.
An AI does not take your word for anything. It decides you are trustworthy by checking whether what you say on your own site matches what everyone else says about you, and where those disagree, it treats the claim as unverified and moves on.
A few weeks ago I put my own website through a free tool that asks four AI platforms the same question about a business, three times each, and scores what comes back on checkable facts against unsupported impressions.
The results were not close.
Perplexity returned fourteen checkable facts about Wisepreneurs and scored +28. It named the Positioning and Practice Audit at AUD $1,950, the marketing partnership from AUD $4,950 a month, the twenty-seven years, the specific audience. It read the live site and played the current positioning back to me almost word for word.
ChatGPT returned one fact and scored +2.
Gemini returned no facts, one unsupported impression, and scored minus 2.
Claude returned no facts, two unsupported impressions, and scored minus 4.
Total, +24.
The tool calls that a solid foundation, which is a generous reading of a result where two of the four platforms could not state a single verifiable thing about a business with twenty-seven years behind it, a hundred-plus published articles and more than ninety podcast episodes.
What was more useful than the number was the pattern underneath it, because the pattern is not really about my website at all.
The four tools disagree because they are doing different jobs
Perplexity was built to search the live web. When asked about my business it went and read the current pages, so it reported the current facts.
ChatGPT, Gemini and Claude lean much harder on what they absorbed during training. Asked the same question, they answered from an older version of me, or from nothing much at all.
ChatGPT reached for vocabulary I stopped using in 2024. Gemini filled the gap politely with words like "solopreneurs" and "coaches", neither of which appears anywhere in my current positioning. Claude had no specific knowledge of the business and gave honest generic advice instead.
That third case is the interesting one, and it is the one most experienced professionals will hit. Claude was not wrong about my site. It had simply never encountered enough independent evidence of my existence to say anything definite.
You cannot fix that by editing a page. You cannot go back and rewrite a training notes for the AI. What you can do is change what the web says about you from here, and make the current version of your practice unmistakable everywhere a machine might look.
Which reframes the job. For twenty years the question was whether you could be found. The question now is whether, once found, you can be corroborated.
A note on the scoring. The tool separates "facts", meaning specific checkable claims, from "vibes", meaning impressions with nothing behind them. A confident, warm, entirely generic description of your field scores worse than nothing, because it means the machine is filling a gap rather than reading a record.
More than half the traffic to your site is already automated
Imperva's 2026 report puts automated traffic at 53% of all web traffic in 2025, up from 51% the year before, with human activity down to 47%. That is the second consecutive year machines have outnumbered people on the web.
Some of that is old-fashioned scraping and attack traffic. A growing share is AI crawlers building an account of who you are, and agents fetching pages on behalf of a person who has asked a question.
Here is what that does to the shape of the problem. In the search era, your site competed for attention with other sites and the reader chose.
In the current arrangement, something reads your site, reads everything else about you, forms a view, and reports that view to a person who may never visit you at all. You are no longer competing for a click. You are supplying evidence to a summariser.
Gill Andrews, a conversion copywriter I interviewed who works almost entirely on the websites of service providers, made a point in that conversation that lands differently now than when she said it.
Design, actually, that's the first thing people notice. Which is ironic because when I work on a website, I always start with words.
A human sees your site design first and your words second. A machine reads only the words, the structure and the markup. Same page, two entirely different readers, and only one of them is now in the majority.
The four questions an agent works through
Liz Micik has spent nearly thirty years on the structural side of websites, the part nobody sees. Platform migrations, domain changes, companies splitting one site into two or merging three into one, on everything from a five-page site to large enterprise sites. Her description of her own work is that she is the person companies call when the structure of the website itself is the problem.
She puts four questions to any site.
- Can an agent find you
- understand you
- trust you, and
- use you.
This is a very useful frame to apply to your website, because it sorts a sprawling problem into four sequential gates, and failing an early one makes the later ones irrelevant.
The four questions come from Liz Micik, Wisepreneurs Podcast episode 96.
Can it find you?
This is access, and it is the cheapest layer to get right.
Your robots.txt file decides which crawlers are welcome. Most professionals have never opened theirs. Mine allows the live citation crawlers, which is what lets ChatGPT and Perplexity fetch a page when someone asks about me right now, and that turned out to matter more than any training-data question.
The other half of finding is whether your content exists in the raw HTML at all. Page builders love to hide content behind JavaScript toggles, tabs and accordions. A person clicks the tab and sees the text. A crawler often does not.
Kate Toon, who has run an SEO and copywriting practice for sixteen years, described a case in our conversation that still makes me wince:
I had somebody who had a website up for an entire year with that clicked. And so all their efforts, all their blog posts, everything had not worked.
Oops!
WordPress has a setting under Settings → Reading called
"Discourage search engines from indexing this site."
It's there so developers can hide a site while they build it, and it's meant to be switched off at launch.
If it's left on, the site works perfectly for anyone who has the link, but Google won't index a single page, so nothing you publish can ever be found in search.
Kate's client had it ticked for a year.
Every blog post they wrote was invisible to search the whole time, with no error message to tell them.
Can it understand you
This is parsing, and it is where the interesting failures live.
Structured data, or schema, is the layer of code that tells a machine what kind of thing a page is and who is behind it.
It has been unfashionable among search specialists for years, because the visible benefits it used to deliver in Google results have largely gone.
Liz's argument, which I think is right, is that the reason to use it has changed rather than disappeared. It will not lift your rankings. It will help an AI resolve who you are and connect the version of you on your website to the version on LinkedIn and in the podcast directories.
The rule that still governs it is the old one: anything asserted in your schema must also appear visibly on the page. Schema that contradicts the page is worse than no schema.
Peter Hatherley, who founded an AI company in Christchurch and has spent years on the semantic side of this, gave the clearest plain-language explanation of the underlying mechanism I have heard:
Say if you think of a kitchen, a cupboard is an entity. It's connected to the kitchen. Spoons are. The sink is. Water is. And so all of those words are connected.
A machine works out what you are about by the company your words keep. If your website says positioning, your LinkedIn says branding, your Google listing says marketing services and an old directory entry says coaching, you have handed it four weakly connected clusters instead of one strong one. It will pick whichever is best supported, and that is frequently not the one you would choose.
Can it trust you
This is corroboration, and it is the gate almost everyone underestimates, including me.
An AI does not take your word for anything. It decides whether a claim is a fact by checking whether the claim appears consistently across independent sources. Where the sources agree, it can state something. Where they disagree, it hedges or drops the claim.
Sue Ellson, an independent LinkedIn specialist in Melbourne on the Wisepreneurs Podcast episode 35, described the human version of this years before AI agents existed, and it has aged into a technical specification:
I believe it's not what you know, it's not who you know. It's who refers you and how you can be verified.
She went further in that conversation, on what happens after a referral.
If you say I've spoken to two international LinkedIn people, but Sue's based in Australia, I recommend you speak to Sue and then somebody Googles me and everything matches what you said about me, then the person's going to book me. They're not going to believe you even if you were their best friend. They will still Google you.
Everything matches what you said about me. That is now a machine-executed test, run in a fraction of a second, on every source it can reach.
Robert Vlach, who has spent more than twenty years working with the Czech freelance community, draws the distinction I keep returning to when people ask whether they should be doing more self-promotion.
A good name, he says, "is mostly created by other people talking about you behind your back". A personal brand is something else, and he is blunt about its failure mode: "You may be known, but not renowned."
The machine-era version of that failure is being visible and uncorroborated, which reads to an agent as a claim rather than a fact.
Can it use you
This is action, and it is the newest and least settled of the four.
An agent that can find, understand and trust you may then try to do something: book a call, submit an enquiry, request a quote, buy a document.
Liz is direct about what stops it.
- Contact forms built so an agent cannot read or submit them.
- CAPTCHAs an agent cannot pass.
- Product details trapped inside a PDF, which she describes flatly: an agent cannot read a PDF, so it does not know what you offer.
Her example is a small manufacturer with three products in many sizes and specifications, sold through a request-for-quote form. The agent has to find every specification, add them up and complete that form. If it cannot, the enquiry does not happen and nobody ever knows it did not.
For an independent professional this sounds remote until you notice that the same logic applies to a calendar link and a contact form, which is the entire commercial surface of most consulting practices. You may have done everything else right and still be blocking the last step.
The question | What it is really testing | Where practices fail |
|---|---|---|
Can it find you | Crawler access and raw HTML | Blocked crawlers, content hidden behind JavaScript, a stray no-index setting |
Can it understand you | Structure, schema, consistent language | No schema, schema contradicting the page, four different descriptions of the same work |
Can it trust you | Corroboration across independent sources | Claims only you make, stale bios, third-party profiles that disagree |
Can it use you | Whether an agent can complete an action | Unreadable forms, CAPTCHAs, specifications locked in PDFs |
The machine layer is the version that gets quoted
Here is the part that surprised me most when I audited my own site properly.
There is a layer of your website most people never look at: the meta tags, the Open Graph descriptions that generate your social previews, the alt text on your images, the schema, and increasingly a plain-text file at your site root that summarises your business for AI tools.
That layer is often years behind your visible copy, because when you rewrite a page you rewrite what you can see.
An agent reads the tags. A person reads the prose.
My own case was instructive and slightly embarrassing. My plain-text summary file listed the marketing partnership at $2,750 a month. The homepage said from $4,950. That file is exactly the sort of thing an AI tool reads first and quotes verbatim, which means a prospect could have been told a price at a little over half the real one, by a machine, with confidence, and I would never have known it happened.
The same audit found the homepage carrying two different descriptions of who I serve. The current one at the top, an older block further down describing people transitioning into self-employment, which is a different audience entirely. An AI extracting that page gets two contradictory definitions and has to choose.
While preparing this article I had the current site checked again, and found more.
- The newsletter page's title tag calls it The Wisepreneur Newsletter while the page copy calls it The Wisepreneurs Insider newsletter.
- The Insight Hub claims 145 articles where the sitemap lists about 105.
Two different category URL sets are in circulation, one in the navigation and one in the sitemap.
None of it is dishonest. All of it is what happens when you publish for years and nobody ever checks across the whole set. And every instance is a small invitation for a machine to decide it cannot verify you.
Consistency is the work, and it is not interesting
My separate article on what your website should do for a Wisepreneur makes the case for the site as the place a prospect works out whether you can solve their problem. This is the machine-side counterpart, and the demand it makes is different. A human forgives inconsistency. They read past it, or never notice. A machine treats it as a reason to doubt you.
So I went looking through my own presence, expecting it to be tidy.
I have been stating twenty-seven years in one place and something different in another. Then I checked a client whose two websites I built myself and found her qualifications listed differently in four or five places across her online presence. Nothing false anywhere. Just a body of work that accumulated faster than anyone audited it.
This is the least glamorous work in an independent practice and I suspect that is precisely why it goes undone. There is no launch. Nobody congratulates you. It shows up, if it shows up at all, as a prospect who arrives already convinced instead of an enquiry that never comes.
Sue Ellson supplied the other half of this argument in a later conversation, after discovering Bing had de-indexed her website entirely:
The tech will discriminate against you, wipe you out... But the good news is if you do still search for me on Bing, all this other content that I've had published about me still appears. So it's not as if I vanished off Bing, but just my website listing has vanished.
A distributed, consistent presence survives the failure of any single property. That is the practical argument for putting the same story in many places rather than perfecting one page. It is also why podcast guesting, guest articles and association work matter more than they appear to: they are the independent sources the corroboration test needs.
What none of this changes
It would be easy to read all of the above as an argument that your practice is now a technical compliance exercise. It is not, and the people furthest into this work are the clearest about that.
Liz's own answer, when I asked whether AI would cut professionals out, went straight to Google's long-standing framing of experience, expertise, authority and trust, and to the specific thing a person brings that a model cannot manufacture.
She hand-coded her first website in 1997. She now runs a one-person practice with an AI assistant she has given a name and a memory of her business, and skills she built herself for the forty checkpoints in her audit.
She still checks every result by hand, and she can do that only because three decades of pattern recognition tell her when an answer looks wrong.
Robert Vlach, who is not an AI sceptic, put the limit plainly: a large group of top-level experts use AI for support work while remaining wary of it supplementing their core expertise, and he counts himself among them.
Debbie Jenkins, who has written seventeen books, described the market problem underneath all of this:
There is a really low barrier to entry from becoming an expert in a company to then be taking that expertise and becoming a consultant or a coach... anybody can claim that credibility.
That is the real reason the corroboration layer matters. When anyone can claim expertise and AI can generate a plausible account of anything, the scarce commodity is verifiable evidence that a specific person has actually done the work. Making yourself legible to machines is not a substitute for having something to be legible about. It is how the thing you have gets counted.
And the work runs both ways. Everything that makes your practice easier for an agent to verify, clearer claims, specific evidence, current information, consistent description, makes it easier for a person to verify too.
Gill Andrews put the mechanism in one line:
trust is achieved by specificity
Vagueness gives a human nothing to believe and a machine nothing to repeat.
How you would know it is working
Honestly, you would not, not yet, and anyone who tells you otherwise is selling something.
There are no established benchmarks for this. Nobody can tell you what a good score is, or what the acceptable share of AI-sourced enquiries should be for a consulting practice in your field.
The metrics people are reaching for, citation counts and appearances in AI summaries, measure being mentioned rather than being chosen, and those are different things.
What you can do is watch the numbers you already own. If bot traffic climbs while your enquiries hold or improve, you are being read and reported accurately. If a growing share of your first conversations begin with the prospect already knowing what you do, the corroboration layer is working.
The cheapest instrument available is a single open text field on your contact and booking forms asking how the person heard about you. It costs nothing, it needs no tool, and it will tell you about a shift in where your work comes from months before analytics does.
Next steps
In order, cheapest and most revealing first.
- Run your own site through a free AI check and read the list underneath the score, not the score. The list is where you see what each platform is getting wrong about you or reporting from three years ago, see Liz Micik below.
- Open your own machine layer. Look at the meta description, the Open Graph tags and any plain-text summary file at your site root, and check every fact in them against what your site says today. Prices and dates first
- Write down the one-line description of what you do, then compare it, word for word, against your LinkedIn headline, your Google listing, your podcast bios, your directory entries and your association profiles. Fix them to match rather than improving each one separately
- Check your contact form and calendar link actually work for something that is not a human. If your form uses a CAPTCHA or a script an agent cannot complete, that is the last gate closing
- Get your specifications, prices and service descriptions out of PDFs and onto pages
- Then, and only then, worry about schema. It is worth doing and it is not where the failures are
None of this needs a rebuild. Most of it is housekeeping. The site should have passed anyway.
Frequently asked questions
Why do ChatGPT and Claude know so little about me when my site has been up for years?
Because those tools lean heavily on training data, which is a snapshot of what the wider web said about you at a point in time, not a reading of your current site. If most of the evidence about your practice lives only on your own site, there was little for them to absorb. The fix is off-site: podcast appearances, guest articles, directory listings and mentions by other people.
How do independent professionals get AI tools to describe them accurately?
Start by making the current description of your work identical everywhere you appear, then add independent sources that repeat it. AI systems treat agreement across sources as evidence and disagreement as a reason to hedge, so exact repetition does more than clever wording.
Is schema markup still worth adding in 2026?
It will not improve your Google rankings, which is why many search specialists dismiss it. It does help AI systems resolve who you are and connect your website to your other profiles, which is a different job and a more useful one now. Add it, but only after your visible facts are consistent, because schema that contradicts your page does harm.
I only have a five-page website. Does any of this apply to me?
More than it does to a large site, because a small site is entirely fixable in an afternoon and because a solo practice has nowhere else for the evidence to live. The larger job for a small practice is off-site: making sure the handful of places you appear all tell the same story.
References
Liz Micik AI Website Visibility: What Independent Professionals Must Fix Now. Wisepreneurs Project podcast, Episode 96, on the four questions an AI agent asks of a website, why structured data matters again for a new reason, and what stops an agent completing an action.
Sue Ellson LinkedIn Specialist and Gigster: From Corporate Banking to Independent Career. Wisepreneurs Project podcast, Episode 35, on referral plus verifiability as the actual mechanism behind getting hired.
A Realistic Guide to Starting a Business After 60 with Sue Ellson. Wisepreneurs Project podcast, Episode 72, on being de-indexed by a search engine and why a distributed presence survives it.
Robert Vlach Professional Reputation: Why Your Good Name Beats Personal Branding. Wisepreneurs Project podcast, Episode 75, on a good name built by other people, and the difference between known and renowned.
Kate Toon SEO and Copywriting: Building an Independent Online Presence as a Solopreneur. Wisepreneurs Project podcast, Episode 47, on the website as validation rather than lead generation, and a site invisible to search for a year.
Gill Andrews Conversion Copywriting: Websites That Generate Inquiries for Independent Solopreneurs. Wisepreneurs Project podcast, Episode 46, on specificity as the trust mechanism, and why design is read first by humans and not at all by machines.
Peter Hatherley Semantic SEO and AI Tools: How Authored Intelligence Serves Independent Professionals. Wisepreneurs Project podcast, Episode 54, on entities and how a machine works out what a page is genuinely about.
Debbie Jenkins Writing Short Books for Credibility: Why Concise Books Build Professional Authority. Wisepreneurs Project podcast, Episode 52, on credibility clues and the collapsed barrier to claiming expertise.
Imperva - Bad Bot Report 2026: Bots in the Agentic Age. Reports that automated traffic reached 53% of all web traffic in 2025, up from 51% in 2024, with human activity falling to 47%.
Liz Micik - Signal Check, lizmicik.com/signal-check. The free diagnostic used for the results reported in this article, which queries ChatGPT, Perplexity, Claude and Google Gemini three times each and scores the responses on verifiable facts against hedges.
Standing close
Every Tuesday I send The Wisepreneur, a short letter for experienced professionals building independent practice: one usable idea a week. Sign up below in the footer.
When you want a second pair of eyes on your own practice, the Positioning and Practice Audit is how I work with people: wisepreneurs.com.au/marketing-partnership