AI Search · Google · Small Business
Google's New AI Search Ranking Model. Why I'm Not Worried.
Google DeepMind put out a paper on an AI that would rank search results by itself. One model doing the whole job. It is research, not a launch. Nothing about your site changed this week. But the direction is clear enough to plan around. Here is what the paper says. What it would do to your phone. And what I am doing about it.
The Short Answer
Google DeepMind proposed a system called Autoregressive Ranking. ARR for short. They worked on it with researchers from the University of Massachusetts Amherst and the University of Texas at Austin.
ARR uses one AI to write the whole ranked list. Search does that job in two passes today. ARR does it in one.
In testing it tied the stronger of the two passes. It beat the weaker one easily. That is it. Google has not said it will ship. Nothing in live search moved because of it.
Source: Search Engine Journal, Roger Montti, September 10, 2026.
3 Things to Remember
One model, one list. No first pass, no second pass.
Google Just Published a Plan to Let One AI Rank Every Page in Search.
The paper is called "Autoregressive Ranking: Bridging the Gap Between Dual and Cross Encoders."
It comes out of Google DeepMind, the University of Massachusetts Amherst, and the University of Texas at Austin.
Here is what they want to do. One AI writes the ranked list by itself. No first pass. No second pass. One model, one list.
That is a bigger swing than it sounds.
Search Engine Journal called it a radical change. They said the hit to SEO and answer engine optimization would be huge if Google ever shipped it. I agree with both halves of that. Including the if.
The machine is quick and dumb. The person is slow and right.
Here Is How Google Ranks Pages Today, in Plain English.
Google ranks your page in two passes. Two machines. Two different jobs.
Pass one is the sorter. Google calls it a Dual Encoder. It turns your page into a bunch of numbers, kind of like a fingerprint. Then it grabs the pages whose numbers sit closest to what somebody typed. Fast. Cheap. Dumb as a rock.
Pass two is the reader. That one is a Cross Encoder. It takes the short pile and looks at the search and your page side by side. Way more accurate. Way more expensive. So Google never points it at the whole web. It only sorts the finalists.
Think of the back room at the post office. A machine slings envelopes into bins off one stamped mark. Then a guy at a desk opens the small stack that made it through.
The machine is quick and stupid. The guy is slow and right.
Google runs both because neither one can do the job alone.
A ranker that reads can tell answering from mentioning.
Google's New AI Search Ranking Model Reads Your Page, Not a Fingerprint of It.
ARR throws out the fingerprint pass. One model does the whole thing and writes the ordered list straight out.
To teach it, the researchers built a training method called SToICaL. That stands for Simple Token-Item Calibrated Loss. In plain words it does two things.
Pages that should rank higher get more weight while the model learns.
And the model gets pushed toward picks that lead to better pages.
What comes out is a model that learns which pages are worth something. And holds the junk down.
They proved one more thing on paper. A Dual Encoder fingerprint has to keep getting bigger as the pile of pages grows. It needs the room to keep every order straight. ARR has no such ceiling. In theory it ranks any number of pages at one fixed size.
That word theory is doing a lot of work.
The researchers say so themselves. A proof on paper is not proof of how it behaves in live search. Google handles billions of searches a day. That is a different animal.
Here is the part that hits your site.
A ranker that reads knows a page that answers the question from a page that just says the words. A fingerprint matcher has a much harder time with that. That is why weak headings quietly cost you in AI search already.
The Test Results Are Good, Not Perfect, and That Matters.
They tested ARR on two data sets. WordNet and ESCI Shopping Queries.
On WordNet, ARR did about as well as the Cross Encoder, the expensive accurate one. And a lot better than the Dual Encoder. It also cut down on junk pages floating above good ones.
Then there is the miss.
On the shopping test, one version got worse at putting the single best result first. The overall order improved. The top slot got worse.
Read that twice.
In shopping search the top slot is the whole game. Sort the list better but put the wrong thing on top and you are not ready for real customers. The researchers flagged it themselves as needing more work.
That is a real gap. It is the kind of gap that keeps a paper a paper.
Being in good shape is not the same as being safe.
I Think We Are in Good Shape on Google's AI Search Ranking Model.
Start with the number that already moved.
68.01% of US Google searches ended with no click at all in the first four months of 2026. It was 60.45% in 2024. That is SparkToro, 2026, off Similarweb clickstream data.
Ranking and getting the click split apart a while ago.
Pew Research Center saw the same thing from the user side. July 2025, 68,879 real Google searches. People clicked a regular result on 8% of visits when an AI summary showed up. They clicked on 15% when it did not. Only 1% clicked a source inside the summary.
So the click got harder to get. Being the answer is what pays now.
There is a number on that too. Seer Interactive looked at 5.47 million searches and 2.43 billion impressions across 53 brands. April 2026.
Get named inside the AI answer and you pull about 20,743 clicks per million impressions.
Sit on that same page and not get named, you pull about 9,445.
Same page. Same search. Twice the clicks for being the one it names.
Seer is careful about what that proves. So am I. They say flat out they cannot claim the citation causes the lift. Strong sites get cited more in the first place. Fair.
It still tells you which side of that line your money is on.
The mistake: you write your service page for a keyword matcher. The phrase goes in eight times. The real answer sits down in paragraph four. And the page opens with a line about your commitment to quality.
The fix: put the answer in the first two sentences of the page and every section. Then name real things. The city. The price range. The steps. The timeline. That is all answer first writing is.
The payoff: the page wins today under the system Google actually runs. And it still wins if a reader ranks it instead. You are not betting on which one shows up.
I will show my work. We have built 281 JSX AI-interactive sites. Every one ships prerendered HTML with the answer at the top. Because being the one AI picks has been the point since day one.
We also moved our own 1,387 posts off a dying WordPress stack. So the content underneath would be readable instead of glued together at load time.
That was not a bet on this paper. It was a bet on the direction. This paper points the same way. The 100K Website is built to survive this kind of shakeup on purpose.
Being in good shape is not the same as being safe.
Nobody outside Google knows the signals of a system that never shipped.
Week one and month four do not get the same year.
I Set Up a Monitor So We Catch This the Week It Moves.
I can't promise I will know before Google announces. Nobody can. What I can do is stop finding out late.
I have a task that runs on a schedule and watches for new work on this. Follow ups from the DeepMind team. Anything that turns the paper into a product. The search press that covers it. If ARR or something like it goes from research to live search, it hits my desk that week.
That is the whole value. Smaller than "we called it," and a lot more useful.
In a search change, week one and month four are not the same year. Month four is when you finally see the drop in your own numbers. By then the fix takes a quarter, not a weekend.
Plenty of owners learned that the hard way when old posts stopped earning AI citations.
Try it · Score one page for a ranker that actually reads it
Paste one real page. You'll see what a reading model sees.
You are a search ranking model that reads a whole page before deciding where it belongs. Here is one page from my website: [PASTE YOUR PAGE COPY] I run a [TYPE OF BUSINESS] in [YOUR CITY]. Score this page from 0 to 100 on how easily a reader like you can tell what it answers, who it is for, and why it should rank above the other pages on the same question. Then list the 5 biggest problems, worst first, in plain words. For each one, give me the exact fix in one sentence. Rules: 5th grade reading level. Short sentences. Plain words only. End with the single change that would move the score the most.
Fill the [brackets] in the chat box before you send.
Prompt · Build your own search change watch list
You don't need my monitor. You can run one.
Act as a search news analyst for a small business owner. I run a [TYPE OF BUSINESS] in [YOUR CITY], and my website brings in about [ROUGH PERCENT] of my new customers. Build me a watch list of the 5 changes in Google and AI search that would actually cost me customers if they shipped. For each one, tell me the first warning sign I would see in my own numbers, and the one place I should check each month to find out if it happened. Name real tools and real websites. Keep it to one page. 5th grade reading level. Short sentences. Plain words only.
Fill the [brackets] in the chat box before you send.
Where I Could Be Wrong About This.
Google puts out a lot of research that never ships. This could be one of them. Search Engine Journal's read is fair. The current encoders still do the ranking work today. We have not hit the point where everything changed.
Plenty of people will tell you otherwise this month. They are ahead of the evidence.
There is also better news in the data than the scary numbers suggest. You should hear it.
Seer Interactive found AI Overviews on about 36% of informational searches. But only about 5% of the ones where people buy.
Rand Fishkin at SparkToro says the same thing from the other side. Local businesses, branded searches, and high intent buying searches are still where search sends real traffic.
Somebody types "emergency plumber open now." Google shows a map, not an essay. If most of your money comes off searches like that, you are less exposed than a publisher is.
Now here is where I could be wrong.
"Clear content wins" is my read of what a reading ranker would reward. It is not a documented ranking factor. There is no documentation for a system that does not exist.
A ranker like this could weigh things nobody outside Google can see. Some of those things might be out of your hands entirely.
So here is the line I hold.
Anybody selling you a fix for a system that has not shipped is selling you a guess. With a price on it. That includes me, if I did it. So I won't.
What I will do is keep building pages that win under the system Google runs right now. And tell you the week that changes.
What To Do Next
You just read a lot of theory about a system that is not live yet. The short version is this. The pages that live through a ranking change are the pages that already answer well.
Get My AI SEO Score. See how your pages read right now, one by one, under the AI search tools already running today.
No contract, no cost to look.
FAQ
No. Autoregressive Ranking is a research paper from Google DeepMind and two universities. It is not a live product. Google has not announced any plan to put it into search. Your rankings did not move because of it.
Autoregressive Ranking, or ARR, is a proposed system. One AI writes the whole ranked list of search results by itself. It would replace the two pass setup search uses today. Today a fast machine picks the candidates and a slower one sorts them. In testing, ARR matched the accuracy of the slower one.
No. Not because of this paper. Do your pages answer the question in the first two sentences and name real things? Then you are already pointed the right way. If they open with a paragraph about your commitment to quality, fix that today. That has nothing to do with DeepMind.
A Dual Encoder turns searches and pages into number fingerprints and grabs the closest matches. Fast and cheap, but not precise. A Cross Encoder looks at the search and the page side by side. Much more accurate, but too expensive to run across the whole web. Search uses the first one to build a short list and the second one to rank it.
Less than it used to. SparkToro's 2026 study found 68.01% of US Google searches ended with no click in the first four months of 2026. It was 60.45% in 2024. Local searches and searches where people are ready to buy hold up better than the rest. That is good news if you run a service business.
Not directly. This is Google DeepMind research about Google's own ranking back end. The other AI search tools run their own systems. But the pattern is the same across all of them. Models that read your content reward pages written to be read.
Watch the Google Search Central blog and the Google Research blog for an announcement. Watch the search trade press for coverage. Or run the watch list prompt above and check it monthly. I run a monitor on this and post when it moves.
Sources cited in this post
- Search Engine Journal, Roger Montti, September 10, 2026. "Google DeepMind Develops New AI Search Ranking Model." searchenginejournal.com
- SparkToro, Rand Fishkin, June 8, 2026. "In 2026, Less than One Third of Google Searches Still Send a Click." Data from Similarweb's US desktop and mobile panel, January to April 2026. sparktoro.com
- Pew Research Center, Athena Chapekis and Anna Lieb, July 22, 2025. Browsing data from 900 US adults, 68,879 Google searches, March 2025. pewresearch.org
- Seer Interactive R&D, April 2026. "AIO Impact on Google CTR: 2026 Update." 53 brands, 5,471,127 queries, 2.43 billion organic impressions, January 2025 to February 2026. seerinteractive.com
Check Out My Last 3 Builds
Real sites, built with this exact system. Tap any one and poke around.