Where did the visitors go
Lately, I've been seeing a familiar kind of distress signal on Reddit. Website owners are looking at their organic search results and wondering what happened. The pages are still there. The business is still there. Somebody is still paying for the website. The visitors, however, appear to have slipped out through the kitchen.
I've been expecting this shift for months. AI is changing where people encounter answers, and the old agreement between search engines, websites and readers is getting rewritten while everybody is still sitting at the table.
I have an impolite name for the manipulation side of search engine optimization: fool the tool.
Figure out what the algorithm appears to reward. Adjust the page. Watch the rankings. Adjust it again. Somewhere along the way, a perfectly ordinary business owner could find themselves discussing keyword density with the seriousness normally reserved for aircraft maintenance.
Meanwhile, the customer wanted to know whether the shop was open on Sunday.
Good SEO professionals have always done useful work: making sites accessible, organizing information and helping businesses answer questions. Google explicitly treats keyword stuffing and link spam as violations of its policies. The manipulation industry has been arguing with the referee for years. Google's spam policies
My objection is to the point at which serving the algorithm becomes the business and serving the reader becomes an optional extra.
Now an AI answer can arrive before that reader ever reaches your carefully arranged page. The front door has acquired a receptionist, and the receptionist has opinions.
There is evidence behind the unease. In a study of browsing data from 900 U.S. adults, Pew Research Center examined 68,879 unique Google searches from March 2025, reconstructing their results in April. Visits associated with an AI summary led to a click on a traditional search result 8% of the time, compared with 15% for visits without a summary. Links inside the summaries received clicks in just 1% of visits with a summary. Pew Research Center
That is a dated observational study, not a diagnosis of every website's traffic problem. It does illustrate why a publisher can remain visible while receiving fewer opportunities to welcome someone through the door.
The index is still here
Google has not hauled its index to the curb with a FREE sign taped to it. Its documentation says a page must be indexed and eligible for a search snippet to appear as a supporting link in AI Overviews or AI Mode. Existing SEO fundamentals still apply; no special AI markup is required. Google's guidance for AI search features
Keep the crawlable pages. Keep the useful links. Keep the sensible headings. They help people and machines understand what is there.
But consider the additional job your content may now be doing. An answer system may retrieve it, select a passage and use it to explain something to a person who has never heard of your business. Your page is being considered as evidence.
This would be an unfortunate moment for its main contribution to be that your company is a leading provider of innovative solutions.
Leading whom? Providing what? Innovative compared with which previous arrangement of adjectives?
Marketing has its place. It becomes much more useful when it has something specific to say.
Answer Integrity
At BuellEA, we're making Answer Integrity the foundation of our AEO approach. AEO means answer engine optimization: helping useful information become discoverable and usable in answer experiences.
Our standard is straightforward: clear, accurate answers, supported by evidence and current where it matters.
Accuracy has always mattered. We haven't invented truth, although I expect someone is already preparing a subscription plan for it.
We are making the responsibility explicit. Identify the question. Support the answer. Explain the conditions that change it. Name someone who checks it. Keep changing facts current. Then make the information accessible enough to be found and understood.
Consider a hypothetical tackle shop. Its website could describe its exceptional selection, unparalleled expertise and commitment to angling excellence. A visiting fisherman could finish the paragraph and still have no idea whether the place sells bait.
Or the page could explain which species the shop supports, what it stocks, when it opens and who to call for today's availability. A guide could distinguish firsthand experience from manufacturer specifications. A conditions report could show when an observation was made and avoid dressing up a forecast as something that already happened.
Those details give a customer something useful to act on. They also give an answer system specific claims it can retrieve and compare. Neither reader benefits from another adjective wearing a necktie.
When AI checks AI
Of course, this is the point where somebody reasonably mentions hallucinations.
Yes. Technology has given us machines that can produce a beautifully organized explanation of something that never happened. Apparently, the software industry looked at human meetings and decided to automate the most troubling part.
Fluent language is not proof. Neither is agreement between two models. They may share the same bad source or make the same plausible mistake.
But model-assisted verification is real. In research published in 2024, Google DeepMind described SAFE, a system that breaks a response into individual factual claims, searches for evidence and assesses whether that evidence supports them. The checking process goes beyond asking the original answer to admire itself in a mirror. DeepMind's factuality research
A system can also produce a better answer without replacing its search index or retraining its answering model. Retrieval-augmented generation combines learned model capabilities with information retrieved while answering. Better evidence or better source selection can change the result. The original retrieval-augmented generation paper
Think of a librarian checking another book before answering. The library building has not moved. The catalog may be unchanged. The answer has improved because the librarian did more useful work.
That does not guarantee every false answer will disappear. It gives us a practical reason to publish material that can survive checking. I would rather build around evidence than bet a client's reputation on the hope that nobody asks a follow-up question.
Start with real questions
For a business moving toward AEO, the first step can be modest. Pick the questions customers actually ask. Find the pages that should answer them. Check the material claims, remove unsupported promises, explain relevant limits and assign responsibility for updates. Preserve what already works.
Then measure what happens. Check answer quality. Observe independent citations. Track referrals and meaningful customer actions. Each tells you something different; a mention in an AI answer does not pay an invoice by itself.
Bring receipts
That is why our approach starts with Answer Integrity. It gives the work a standard we can inspect and maintain, even as providers change their models and rearrange their search pages.
I still expect algorithms to change. I've worked in technology long enough to be suspicious of anything described as the final version.
But I want a client's website to be useful when the next change arrives. I want its answers to hold up when a customer, a competitor or another model checks them.
The next question might be the best thing to happen to a business that knows its subject.
It might be as simple as: Can you show your work?
Bring receipts.