AI is not replacing you. It is ignoring you.

A story that made the rounds this year: a New York real-estate broker was closing a fifty-million-dollar penthouse. The buyer asked ChatGPT whether the price was too high. It said yes, and the buyer moved to pull out. The seller asked the same assistant the same day. It told them to hold out for more. Two people, one chatbot, a deal nearly lost on both sides at once.

The point is not the anecdote. It is that the same conversation happens before anyone reaches your website. Buyers ask an assistant who to hire, what to buy, which company does this in their country. The answer is not ten links. It is three names, sometimes five. If yours is not among them, the sale is over before your homepage loads.

The numbers people are quoting

We do not have industry-wide measurements of our own yet, so treat these as what they are: figures reported by operators who track their own funnels. A podcast host who ran a months-long optimisation programme published that visitors arriving from ChatGPT converted at roughly nine times the rate of visitors from organic Google on her newsletter, that a third of US shoppers now start inside an AI assistant, and that her show went from absent to the fifth most visible of the eleven brands she tracked. A marketing survey she cites says most marketers plan to optimise for AI search and fewer than half have started.

Whether your own numbers look like that is exactly the question. The honest way to find out is to ask the engines real buyer questions and record what they say, on a date, with sources. That is what a GEO audit is.

Why a perfectly good website can be invisible

The same host had interviewed dozens of well-known technology leaders. When she asked the assistants for podcasts in her category, her show never appeared. The reasons turned out to be boring and fixable:

• The crawlers saw an empty shell. Transcripts and descriptions were rendered by JavaScript after the page loaded. Humans saw a full site; a crawler reading the source saw almost nothing.
• The machine had her filed under the wrong label. Structured databases that models read (Wikidata, the podcast directories) still described her as a video blogger. A category mismatch keeps you out of category answers no matter how good the content is.
• Pages had never reached an index. A sitemap submitted to Search Console surfaced dozens of pages the engines had simply never discovered.

None of these are content problems. They are plumbing problems, and they are the first things we check.

The playbook, in the order it works

1. Find out whether machines can read you at all. Fetch your pages the way a crawler does: no JavaScript, plain source. Check robots.txt for GPTBot, ClaudeBot, PerplexityBot and Google-Extended. If they are blocked or the source is empty, nothing else on this list matters yet.
2. Put the substance in the HTML. Full text, transcripts, product facts, in the page as it is served. Server-rendered pages are the single highest-leverage change most sites can make.
3. Say one identical sentence about yourself everywhere. Wikidata, your directories, your profiles and your homepage should agree word for word on what you are. When sources disagree, models play safe and name someone else.
4. Title pages the way people ask. Models search internally with specific phrases while composing an answer. A page whose title matches the phrase wins the citation; a clever title does not.
5. Measure weekly, judge monthly. Track a fixed set of buyer prompts across engines and read the trend over weeks. Indexing alone took the host about four weeks; a single week tells you nothing.

The part nobody can shortcut

Most of what the engines cite is earned: articles, listicles, forum threads, other people's pages that mention you. Anyone can call themselves the best on their own site, so models weight what others say. The broad, category-wide questions are won there, and no technical fix substitutes for it. What the technical work does is make sure that when someone finally writes about you, the engines can read, classify and quote you correctly.

What we do about it

Our free audit runs steps one to three for one domain and records the evidence: crawler access, what the source actually contains, structured data, indexation, and, where an engine can be queried, real buyer prompts with the date, the engine and who was cited instead of you. Unavailable evidence is marked not observed, never estimated. Monitoring subscriptions repeat the prompts every week or every month so you can see the trend the host describes, on your own domain, with your own numbers.

Start from the homepage. No signup; results appear on a private page.