The Real Limits of AI Search Engines in 2026

I wrote a piece a few years ago arguing that AI was a bad tool for search engines and that it would damage the products that adopted it.

I was partly right. I was mostly wrong, and being wrong in public is more useful than the original take, so here is the honest accounting.

What I Got Right

Confident wrong answers are worse than no answer. This held up completely. A search result that is a ranked list communicates uncertainty. A generated paragraph communicates confidence it has not earned.

Early AI answers got things wrong in ways that were hard for a user to detect, and that is a genuinely different failure mode from a bad link.

Publishers get squeezed. Also correct. When the answer appears in the interface, the click does not happen. That has real consequences for anyone whose business depends on search traffic.

Attribution stays murky. Still true. Citation practices vary by platform and change without notice.

What I Got Wrong

I underestimated how bad traditional search had become. The comparison I was making was against an idealized version of search results that had not existed for years.

For a huge class of queries, ten blue links of SEO-optimized affiliate content was already a worse experience than a synthesized answer. The bar I thought AI had to clear was lower than I claimed.

I assumed the errors would not improve. They improved substantially. Grounding answers in retrieved sources rather than model memory fixed a large share of the confident nonsense.

I thought it would kill SEO. It changed which pages win. The underlying work of being genuinely useful and technically accessible got more important, not less.

Where AI Search Is Still Weak

Anything time sensitive. Retrieval and index freshness lag. For news, prices, and availability, traditional search still wins clearly.

Local intent. Finding a plumber who can come today is a structured data problem with a map, and a generated paragraph is a worse interface for it.

Transactional queries. When somebody wants to buy a specific thing, they want a page with a button. Summarizing three retailers is not helpful.

Deep research where sources matter. A synthesized answer flattens the distinction between a peer reviewed study and a blog post. For serious work you want to see and judge the sources.

What This Means for Content

The practical implications are less dramatic than the discourse suggests.

Be the source that gets cited rather than the page that gets clicked. Those increasingly diverge, and citation is now a legitimate outcome to measure.

Write direct, extractable answers. A clear self contained paragraph answering a specific question is the unit that gets pulled into generated answers.

Keep the pages that serve transactional and local intent, because those queries are not going through a generated answer anyway.

I laid out which of the surrounding advice is real in GEO vs SEO, and what the signals actually mean in what actually matters about AI search.

The Lesson I Take From Being Wrong

I was evaluating a new technology against an idealized version of the incumbent. That is a specific and common analytical error and I have watched a lot of marketers make it since.

The question is never whether the new thing is good. It is whether it is better than what people are actually using, which is usually worse than you remember.

For tracking how this keeps shifting, Google's own documentation on AI features in Search is the only first party source, and it is worth reading over any third party speculation.

Should You Change Your SEO Strategy Because of AI Search?

Adjust it, do not rebuild it. The work that makes a page useful to a person is still the work that makes it eligible for citation. What changes is which pages you prioritize and what you count as a result.

Three concrete adjustments. Write direct, self-contained answers to specific questions, because that is the unit that gets extracted. Keep transactional and local pages strong, since those queries still resolve through traditional search and maps. And start treating citation without a click as a real outcome rather than a failure.

That third one requires a reporting change more than a strategy change. If your success metric is sessions, a channel that produces awareness without visits looks like a decline. Branded search volume and direct traffic to deep pages are the signals that catch it.

What I would not do: rewrite your entire site for machine consumption, buy a tool that promises AI ranking, or chase the advice cycle that reinvents SEO terminology every few months. Most of that is repackaging fundamentals with new names, which I went through in detail in GEO vs SEO.

The honest summary for most businesses is that the highest value work has not changed. Fix the technical basics, answer real questions properly, and earn genuine links. Those inputs feed every discovery surface that exists.

The broader habit I took from getting this wrong: when judging a new technology, write down what people actually do today rather than the idealized version of it. I compared AI answers against search results as they existed in my memory, not as they existed in 2023, and the gap between those two was the whole error.

What did you predict about AI that turned out wrong? Mine was this, and the correction was more useful than the prediction.