Google’s own AI now takes most of its answers from outside Google’s first page. We tested four assistants to find out what they use instead.
For twenty years marketing had one job: be findable. You ranked, you got clicks, you got customers. The work was buying your way up a list — backlinks, domain authority, volume, bought over years before it paid.
That job is ending. The thing replacing it does not run on the same rules, and most of what the old job rewarded is worth close to nothing in the new one.
This page reports what we found when we tested that directly.
Google’s own AI has largely stopped using Google’s rankings
Start with the part that surprises people who assume their ranking protects them.
Ahrefs looked at 863,000 keywords and four million AI Overview citations. In July 2025, 76% of the pages Google’s AI cited also ranked in Google’s own top ten. By March 2026, 38%. About a third of citations now come from pages ranked between 11 and 100, and another third from pages ranked past 100 — pages that, for practical purposes, do not rank at all.
BrightEdge measured it differently and reached the same place: only 16.7% of AI Overview citations come from top-ten results. Their conclusion was blunt — ranking well does not guarantee AI citation.
And that is the generous case, because it is Google citing Google. ChatGPT, Claude and Perplexity do not run on Google’s index at all. A number one ranking is an asset inside one company’s ecosystem. It does not travel.
So we tested what the assistants actually do
In September 2026 we asked four AI assistants — Google’s AI, ChatGPT, Perplexity and Claude — the two questions a homeowners association board asks when it starts looking for community security. Who should guard our community. How do we alert residents in an emergency.
Between them the four assistants named twenty-six companies. We have not named any of them. We are one of the companies that did not appear.
Scale did not decide it. The three largest security firms operating in this market — companies with Wikipedia entries, Bloomberg profiles and pages of local job listings — were named a combined total of once, across eight searches on four engines. Two of the three were never named at all.
One local firm with twenty-one Google reviews was recommended by all four assistants.
What the answers were built on instead
Every assistant justified itself the same unglamorous way: a state licence number, a star rating, a service list, and a page on the company’s own website that plainly said it did this work. On the software question, the assistants quoted specific product detail — one vendor’s pricing down to the dollar, another’s two-way reply feature — all of it lifted straight off the vendors’ own sites.
What never appeared anywhere: news coverage, trade press, review sites, Wikipedia, or any business database. All the assets that take years and budget to accumulate.
Then we asked one of them to show its working
We asked one assistant to disclose the evidence behind each of its recommendations. That disclosure is the most useful thing in this audit.
It had obtained no star ratings and no review counts, because the review platforms block automated access. It had verified no licence numbers — every number it printed came from the vendor’s own page. It had obtained no client references. Its top recommendation was ranked first on the strength of a trade association membership the company reported about itself, on its own About page.
It had also found nine firms in this market carrying independent accreditation and A or A+ ratings — the only third-party-verified companies in its entire research pass — and discarded all nine, because none of them advertised working with homeowners associations.
Asked to summarise its own method, it wrote that its recommendation list was “effectively a ranking of who writes the best website copy.”
We are calling this self-description bias: when an answer engine ranks vendors on the clarity of their own self-description while ignoring independent verification available in the same search results.
What that means if you are the one doing the marketing
Researchers at Princeton and IIT Delhi tested this properly — ten thousand queries, published through ACM. Adding statistics to a page made it 30 to 40% more likely to be quoted. Adding quotations, 41%. Citing outside sources, about 30%. Writing in an authoritative tone: roughly 10%, which they flagged as surprisingly ineffective.
Keyword stuffing produced little to no improvement, and in some cases performed worse than doing nothing at all.
Read that last line again, because it is the whole argument. The core technique of the thing companies have paid agencies for across two decades does nothing here.
Being found rewarded budget. A small company lost to whoever spent more, and lost for years before the spending paid off. Being quoted rewards something else entirely: whether you have published anything specific enough to quote. A firm with twenty-one reviews and a page that plainly says what it does beat companies a thousand times its size — not through a loophole, but because that is the mechanism, and one of the machines will describe it to you if you ask.
If you are on the buying side of this, the same finding is a warning. The list an assistant hands you is not a quality ranking, it is a description ranking. Verify the licence number with the state yourself, ask for references from comparable organisations, and confirm insurance directly — those are the checks the assistant did not run.
Method
Engines: Google AI Mode, ChatGPT, Perplexity and Claude. Two queries, one run per engine per query, conducted 10–11 September 2026. Query one: best HOA security company Las Vegas. Query two: emergency notification system for HOA communities. The Claude run was additionally asked to disclose the evidence behind each recommendation.
What this does not show
One run per engine, on one day. These answers vary between runs — repeat the queries and you will get a different list, and that variability is itself part of the finding. Perplexity required an account for the second query, so that result is unmeasured rather than empty. Two queries is a narrow slice of how anyone might phrase this. Reddit appeared in none of the eight searches, which is worth noting only because so much current advice points there; one engine ran a targeted Reddit search and returned nothing at all.
This audit measures what the assistants said. It does not measure whether any recommendation was good.
Repeat it yourself
The method is simple enough to copy deliberately. Take the two questions your own buyers ask, run them on four assistants, and record which companies get named and what the answer cites underneath. It takes about an hour, and it will tell you more about your visibility than any report you can buy.
We will re-run this quarterly and publish what changes. The last time an assistant reweighted a major source, the shift happened over four days.
Conducted 10–11 September 2026 by VALANTRI, a division of Vigilant Peak Holdings Inc. NV PILB #4427. No company is named in this report. VALANTRI was not named by any of the four assistants.