I ran a technical audit on my own website a few months ago. Not a normal SEO audit. An audit of how legible my site is to machines... to the models answering questions for people who would once have typed those questions into Google and clicked a blue link.
My score came back at 31 out of 100.
I build software for a living. I have written about AI adoption for two years. And my own site was close to invisible to the systems now standing between me and every reader I have.
If you have not checked yours, I have bad news about the number.
The click is going away
SparkToro ran Similarweb clickstream data for the first four months of 2026 and found 68.01% of US Google searches ended with no click at all. In 2024 the same measure was 60.45%. Seven and a half points in two years is the fastest move in a decade of this being tracked.
Pew Research put a microscope on the mechanism. Across 68,879 real searches from 900 US adults, pages with an AI summary sent a click to a search result 8% of the time. Pages without one sent a click 15% of the time. Half the traffic, gone, purely because a summary appeared above it.
The links inside the AI summary itself? Clicked on 1% of visits.
And 26% of sessions ended right there on the summary page, against 16% when no summary appeared. The user got their answer. The visit to your site never happened.
Your analytics will show this as a slow bleed with no obvious cause. Traffic down, rankings unchanged, nothing broken. Nothing is broken. The interface changed underneath you.
Your content strategy assumed a reader with a browser
Every piece of advice you have followed for fifteen years rested on one assumption: a human sees a list of links and picks yours. Titles were written for a human. Meta descriptions sold the click to a human. Page layout was built to hold their attention once they arrived.
Now the first reader of your page is a model. It reads, extracts, summarises and answers. The human reads the answer. Sometimes your name appears next to it. Sometimes it does not.
Two different jobs, and most companies are still doing only the old one well.
Being the top result matters less than being the source worth quoting. Those are not the same skill. A page engineered to keep a visitor scrolling is often terrible source material... long wind-up, claims with no numbers, nothing quotable in a clean sentence.
What the machines reward
I spend most of my time on the engineering side, so this is the part I find genuinely interesting. The retrieval layer does not reward the things your marketing team optimises for.
It rewards content it parses without guessing. Clear headings matching the questions people ask. Facts stated as facts with a number and a date attached. Structured data in the page telling the model what kind of thing this is and who wrote it. Author identity it verifies across more than one place on the internet.
Pew found something revealing about what gets cited. Wikipedia, YouTube and Reddit together made up 15% of sources in AI summaries. Government sites appeared as 6% of sources in summaries against 2% in standard results. The pattern is clear enough. Machines favour sources with obvious structure, obvious authority, and text making a plain claim without a sales pitch wrapped around it.
Your product page with three paragraphs of adjectives before a single concrete statement is not winning here.
What I did about my own 31
I fixed it. The score sits at 57 now. Still not where I want it, and the remaining gap is the hardest kind. Here is what moved the number, in rough order of effort against payoff.
Let the crawlers in. Go and read your own robots.txt. A lot of sites blocked AI user agents in a panic during 2024 and never revisited the decision. Blocking them removes you from the answer, and the answer is where the audience now lives. Decide deliberately, not by leftover default.
Publish a sitemap with real lastmod dates and an llms.txt. Cheap to add, and it hands a model a map of your site instead of asking it to work one out.
Put schema on every page. Article or BlogPosting markup with an author, a date and a publisher. Organisation and Person markup so the entity behind the writing is machine-readable. This is a couple of hours of work with a permanent effect.
Write the first paragraph as the answer. Not a hook. The answer. If someone asks the question your page addresses, the opening lines should stand alone as a correct, complete, quotable reply. Everything else on the page backs it up.
Attach numbers and dates to claims. "Most leaders struggle with this" is unquotable. "68% of searches ended without a click in early 2026" gets lifted verbatim, with a citation. One of these grows your reach. The other disappears.
Name the author and make them real. A byline pointing to a bio page, linked to profiles elsewhere. Models weigh who said it, and an unsigned page from an unknown entity is the easiest thing in the world to ignore.
None of this is exotic. It is a fortnight of unglamorous work sitting in a backlog behind a feature nobody asked for.
The part no engineering sprint fixes
Here is where my own score stalled. Technical work took me from 31 to 57. Everything above 57 needs something no configuration file provides: a name appearing in enough places, said by enough other people, for the model to treat me as an authority worth citing.
You earn it by being quoted elsewhere. By publishing original numbers other people reference. By appearing on other people's platforms. By having a position specific enough to be worth attributing to you.
Which means the work is the same work it always was. Do something worth talking about, and say something worth repeating. The tooling changed. The ranking signal moved from links to citations. The underlying requirement did not shift at all.
The sites disappearing fastest are the ones with nothing original in them... aggregators, rewrites of other people's research, pages built entirely to rank. They have no reason to be cited, because a model reads their source directly and skips them.
Go look at your own number
Pull up your site. Check three things before your next meeting.
Does robots.txt allow AI crawlers. Does your highest-value page state its core claim in the first two sentences. Does any page on your site carry author schema.
If the answer is no three times, you are not competing badly in AI search. You are not in it.
I got a 31 while writing about this stuff for a living. What do you reckon your site scores... and when did you last check?