More than half of all web requests now come from machines.
If your traffic has been sliding and you can’t work out why, this is most of the answer. It’s happening across the whole web, it isn’t your fault, and the shift it’s part of rewards businesses that adapt to it early.
AEOiQ is a free, independent research project on how AI is changing the way things get discovered online. We publish what we find: the data, the methods, and the parts that don’t work yet. No products. No gated downloads. No sales calls.
AI-REFERRED TRAFFIC vs. EVERYTHING ELSE
U.S. retail · 1T+ visits
A year ago
about half the conversion rate of other traffic
Now
54% better
+53% time on site
+23% pages per visit
Your website now has two audiences.
For twenty-five years a website had one job in search: rank high enough that a person clicked through. The page was the destination.
Answer engines changed the shape of that. When someone asks ChatGPT, Claude, Gemini, Perplexity, or Google’s AI Overviews about your category, a model assembles an answer from what it has read. Your site can inform that answer without ever being visited. You can be the reason someone buys and never appear in your own analytics.
So there are two audiences now. People, who still browse and click. And machines, which read, summarize, and decide what to repeat.
Optimizing for the first is search engine optimization. Optimizing for the second is answer engine optimization. They overlap. They are not the same work.
52%
share of AI crawler requests made for model training as of June 2026, up from 22% in spring 2025. Training crawlers read your content and send nobody back.
What actually changes.
Most of what you know about SEO still applies. Crawlable pages, clear structure, real expertise, and a reason to be trusted all still matter. What changes is the unit of success and how long it takes to move.
What you’re competing for
SEOA ranked position
AEOA citation inside an answer
What gets evaluated
SEOThe page
AEOThe passage
How many can win
SEOOne slot per position
AEOSeveral sources per answer
How you know it worked
SEORankings, clicks, sessions
AEOMentions, citations, how you’re described
How fast it moves
SEODays to weeks
AEODays to weeks for retrieval, months for what the model remembers
Where the answer comes from
SEOOne live index
AEOLive retrieval plus what the model already learned
That last row explains most of the confusion. There are two separate channels. A model can look something up while answering you, or it can rely on what it absorbed during training. Fixing your site changes the first within weeks. The second only changes when a model is next trained, and nothing you do makes that happen faster.
Any advice promising to change what an AI “thinks” of your brand next week is talking about one channel only, whether or not it says so.
If your traffic is falling, here’s what’s probably happening.
Search didn’t collapse. It changed shape.
You may have heard that search traffic was going to fall off a cliff. In 2024, Gartner predicted a 25% drop by 2026, and the number went everywhere. Decks, newsletters, sales emails.
2026 is here. It didn’t happen. Overall search volume held up.
Anyone still quoting that figure at you is selling something. But the reason it felt true is that something real did change, and it changed in a way that shows up in your analytics long before anyone explains it to you.
People are getting their answer before they click.
When Google shows an AI summary above the results, people click through 8% of the time. Without one, 15%. Only 1% click a source cited inside the summary itself. Pew tracked roughly 69,000 searches by 900 U.S. adults to get those numbers.
Your traffic can fall while your business is fine. The question got answered on the way.
That is also why the drop feels arbitrary. It isn’t tied to your rankings, your site speed, or anything you changed. It tracks how often your customers’ questions can be answered without leaving the page.
The visits you still get are further along.
Here’s the part nobody leads with. Someone who arrives after an assistant has already answered three of their questions is not a browser. They have compared options, narrowed the field, and come to you on purpose.
That is why AI-referred visitors convert better than any other channel, spend longer, and look at more pages. Fewer people arrive. More of them are ready.
What this means for you.
The goal moved. It used to be how many people find me. Now it’s am I the thing that gets recommended.
That’s a smaller target and a more reachable one. You don’t need to outrank the whole internet. You need to be the answer a model can quote confidently when someone asks about your category, in your area, at your price.
Most of the work is unglamorous and most of it you can do yourself.
Eight things any website can do this week.
None of these need a budget, a vendor, or a rebuild. They are ordered by what they return for the effort.
- 1.
Find out whether you’re blocked.
Open
yoursite.com/robots.txt. Look forGPTBot,ClaudeBot,PerplexityBot,Google-Extended. Plenty of sites block the engines they want to appear in, usually because someone pasted a rule years ago. Decide deliberately. Blocking is a legitimate choice, but it should be a choice. - 2.
Ask the engines about yourself.
Open the assistants your customers use and ask what you’d ask if you’d never heard of you. “Best X for Y.” “Alternatives to [competitor].” “Is [you] any good?” Write down what comes back. Free, twenty minutes, and the single most clarifying thing on this list.
- 3.
Answer one question per heading.
Make headings the actual question a customer asks, then answer it in the first two sentences underneath. Models retrieve passages, not pages. A heading reading “Our Solutions” retrieves nothing. “How much does X cost for a small team?” retrieves cleanly.
- 4.
Put the facts where they can be lifted.
Prices, hours, coverage, specifications, who it’s for and who it isn’t. Write them in plain sentences rather than burying them in an image, a PDF, or an interactive widget. If a person has to click to reveal it, a model probably didn’t read it.
- 5.
Date things.
Recency is a real signal. Add and maintain visible “last updated” dates on pages where the facts change. Undated content of unknown age is easy to skip.
- 6.
Add structured data.
Organization,FAQPage, andProductmarkup describe your facts in a form machines don’t have to interpret. Validators are free. It is the best hour of technical work on this list. - 7.
Fix how you’re described elsewhere.
Models learn from the whole web, not just your site. Review listings, directories, forums, and reference pages often carry more weight about you than your own copy. Correcting a wrong description on a widely-read third-party page can move more than a month of work on your own.
- 8.
Take a baseline you can compare against.
Write down today’s answers to your questions from step 2, with the date. Repeat monthly. You cannot tell whether anything you changed worked without a before.
None of these are one-off fixes. Step 8 is the one that turns the other seven into a practice: once you have a dated baseline, every change you make becomes testable, and a month of guessing turns into a month of evidence. Do step 2 and step 8 this week and you’ll have something in October that almost nobody in your category has. A record of how you were described over time, and proof of which changes moved it.
Want the longer version? Check your AI visibility walks through the same ground with more detail and a worksheet you can keep.
Why this exists.
AEOiQ started with a small experiment. We asked ChatGPT and Claude to recommend products in categories we knew well, then compared the answers to Google. Some well-known brands were missing entirely. Others we had never heard of were described as market leaders.
That gap is the thing this project measures. We track how a fixed set of brands surfaces across five answer engines, week after week, and publish what moves and why.
We’re not selling the fix. Tracking AI visibility is on its way to being a commodity, and the useful thing to do with a commodity is give it away.
What we publish.
Field notes. What we’re seeing in the data. Findings, methods, and the experiments that didn’t work.
Resources. Free guides and frameworks, starting with Check your AI visibility. More coming, including tooling you can run inside your own AI assistant to check your own visibility without paying anyone for the privilege.
Newsletter
We only send something when we’ve found something worth your time.