Eight AI search mistakes, in rough order of cost.

The mistakes in this discipline are consistent enough to list. Almost all of them come from treating AI search as a content problem when it is mostly a data and corroboration problem.

The eight

• Publishing before fixing identity. Content a model cannot attribute to a confidently identified business does very little. Entity work first.
• Measuring with a single ask. Responses vary between sessions. One check tells you almost nothing, and acting on it means acting on noise.
• Blending assistants into one score. ChatGPT, Perplexity and Gemini behave differently and may need different work. An average hides which one you are failing.
• Chasing volume in directories. Two hundred inconsistent listings are worse than twelve accurate ones, because contradictions compound.
• Generating content at scale. Large volumes of adequate AI-drafted content dilute the quality signals that make a site worth citing.
• Ignoring old and defunct listings. A stale address on a forgotten directory still feeds the picture a model builds of you.
• Optimizing informational queries for traffic. Those increasingly earn citation without clicks. Value them as brand presence, not as visits.
• Believing anyone who explains the algorithm. Source selection is not documented. Confidence about internals is a warning sign, not expertise.

Four more mistakes worth adding

The eight above cover the most expensive errors. Four more show up often enough to mention:

• Blocking the wrong crawler. Blocking OAI-SearchBot or PerplexityBot to stop training, when the training crawlers are GPTBot and others, removes you from answers without achieving the goal.
• Ignoring Bing. Copilot and several other assistants rely on Bing's index. A site neglected in Bing Webmaster Tools is neglected in those answers.
• Changing prompts mid-measurement. Editing the prompt set resets the baseline, so progress can no longer be measured.
• Not reading the answers. Teams track whether they are named and miss that the assistant is describing them wrongly.

The pattern underneath them

Six of the eight come from the same root: treating this as a publishing exercise. The instinct is understandable, because content is what agencies are set up to sell and what teams know how to produce.

But the binding constraint for most businesses is not that they have too little content. It is that a model cannot confidently say who they are, or has nothing corroborating them outside their own domain. Neither is fixed by publishing.

How to recover from each

Most of these mistakes are reversible. Content published before entity work can be kept and improved once identity is fixed. Inconsistent directory listings can be corrected or removed. Thin AI-generated pages can be consolidated into fewer, better ones and redirected.

The one that is hardest to recover from is lost history. If measurement was inconsistent or never started, there is no way to recreate it. Start a consistent baseline now, even if everything else is still being fixed.

A quick self-check

Answer yes or no: Do you have a written prompt set that has not changed in three months? Have you checked robots.txt and your firewall for AI crawlers this year? Are your name, address and phone identical on your top ten listings? Do your main service pages answer their question in the first two lines? Have you read, not just scored, what ChatGPT says about you this month?

Each “no” is a place to start, and most of them take less than a day.

Frequently asked questions

What is the biggest mistake in AI search optimization?

Publishing content before fixing entity consistency. A model that cannot confidently identify your business will not cite it regardless of how much you publish, so the content investment is wasted until identity is resolved.

Does publishing more content improve AI visibility?

Usually less than expected, and publishing at volume can hurt. The common binding constraints are identification and corroboration, neither of which is solved by more articles on your own domain.

How many times should I check AI visibility before acting?

Several asks per prompt per period, at minimum five. Responses vary between sessions, so a single check cannot distinguish a real result from normal variance.

Can a bad AI search strategy hurt my SEO?

Yes, if it involves mass-produced thin content or manipulative tactics that fall under Google's spam policies. Accuracy and clarity work helps both.

What should I fix first if I made these mistakes?

Crawler access and entity consistency, then start a consistent measurement baseline. Content fixes come after.