The Machine Isn’t Reading Your Website. It’s Reading What Other People Said About You.

Try this before your next planning meeting. Open ChatGPT or Perplexity, type your company name and the word “reviews,” and read what comes back.

Most of it will be roughly right. Some of it will be positioning you retired two rebrands ago. There will probably be a competitor comparison you have never seen, lifted from a roundup written by somebody who has never opened your product. And if you are unlucky, there is a confident sentence about your pricing that is wrong by a wide margin, delivered in the same calm tone as the parts that are correct.

Now trace it. Click the citations. Almost none of it came from the pages your team spent the last year rewriting. The surface you have been optimizing is not the surface being read.

You are the least trusted source about yourself

Here is the part nobody selling an AEO retainer says out loud. When a model answers a question about a company, that company’s own website is one input among many, and it is the input with the worst credibility profile in the pile. You are the defendant testifying about yourself. Retrieval systems are built to prefer corroboration: the same claim showing up in several places that are not owned by the party making the claim.

So the answer gets assembled from other people. A forum thread. A review site. A comparison article on a domain that outranks you and cares nothing about you. A podcast transcript. A support question somebody posted three years ago that never got a decent reply. Your homepage contributes the spelling of your name and maybe a category label.

Which is why the standard checklist only gets you so far. Schema markup, an FAQ block, a tidy file at the root of your domain telling the crawlers how to behave. That work is not wasted: it controls whether you get parsed correctly and recognized as the right entity at all. Skip it and you are invisible. But it only buys you a seat in the room. It does not decide what gets said about you once you are in there, because the substance is being pulled from somewhere else.

Your brand is a consensus now, not a message

Marketing spent decades on message control. Get the words right, get them on the site, repeat them until they stick. That model assumed the buyer would eventually land on a page you owned.

The buyer now asks a machine, and the machine returns a synthesis. Not your claim. The average of every claim it can find, weighted toward whatever gets repeated by parties with no stake in you. You cannot edit an average by shouting louder on the one channel you control. You edit it by changing what the other sources say.

That is a genuinely uncomfortable shift, because it moves the job from copywriting to something closer to evidence collection, and evidence collection does not photograph well in a board deck.

What actually moves it is boring and slow

None of this is a hack. It is mostly maintenance work that nobody gets promoted for.

• Find where you are described wrong, and fix it at the source. Search your own name across review sites, comparison pages, directories, aggregators, old news coverage. You will find dead product names, a funding round from two rounds ago, a founder who left, a category you no longer sell into. Every one of those is being read as fact. Most of those pages have a “suggest an edit” link and a human behind it.
• Ask customers for specifics, not stars. A five-star review that says “great team” is worth almost nothing to a retrieval system. A review that says what the product replaced, what it costs, who on the team uses it, and what it did not do well is a durable, quotable, corroborating source. Ask for that shape explicitly.
• Answer questions where they are asked, under your real name. Not a content play. Somebody in a community is asking the exact question your buyers ask, and the current best answer is wrong. Go correct it and say who you work for. Anonymous astroturfing gets sniffed out, and it deserves to.
• Publish the numbers only you have. Not a survey of 200 marketers you paid a panel for. The operational reality you can see from inside your own business. Original data is the one thing that gets cited by other people, and citations by other people are the currency here.
• Make pricing legible. If your pricing page says “contact us,” a model will happily fill that gap with a guess from somewhere else, and you will not get a say. A plain page with a plain number is a quotable fact.

Three of those five happen entirely off your domain. The other two live on your site but only pay off when somebody else quotes them: original numbers get cited, and a plain price becomes the fact repeated elsewhere instead of a guess. Either way the payoff is off-site. That is the whole point.

You are not going to get a clean report on this

Here comes the second uncomfortable part. There is no tidy measurement of it.

Some AI referrals arrive with a recognizable source and land in your analytics. Plenty do not. A buyer reads a synthesized answer, forms an opinion, never clicks anything, and shows up eight weeks later typing your name straight into the address bar. That session gets filed as direct traffic. The work that actually did the convincing is invisible to every tool you own, and no amount of tagging fixes it, because there was never a link to tag.

So you run a coarse loop instead. Write down the ten questions your buyers actually ask, the ones that end in a purchase decision. Ask them to the major assistants once a month, same wording. Save the answers and read them side by side over a quarter. You are not measuring a channel. You are watching a reputation drift, and the drift is legible even when the attribution is not.

It is unsatisfying. It will not fill a slide. It is still better than the alternative, which is pretending the surface does not exist because it refuses to produce a chart.

This does not fit in a campaign

The reason most teams will not do any of this is structural. It has no launch date, no creative review, no clean owner. It looks like customer support, PR, and data janitorial work wearing a trench coat. It cannot be handed to an agency as a deliverable, because the deliverable is other people voluntarily saying accurate things about you.

But that is the surface now. A machine is summarizing your company to strangers all day, from sources you did not write and mostly have not read. You can keep polishing the page it barely consults, or you can go find out what it is actually quoting.

Start with the audit. Search yourself, click the citations, and write down every wrong thing you find. It takes an afternoon and it is usually a bad afternoon.

And for the rest of it, the parts that do produce numbers, it helps to only have one set of them. THE DASHBOARD will not tell you what a language model says about your company. Nobody can tell you that reliably yet, and anyone promising otherwise is selling. What it will do is put your marketing, sales, and revenue data in one place so that when you finally go do the unmeasurable work, you are not also refereeing an argument about which tool has the right pipeline number.