How AI Search Actually Works—and Where Marketers Waste Their Time

AI search has a familiar marketing problem: the tactics arrived before the diagnosis.

A new visibility report appears. Someone mentions llms.txt in a meeting. A competitor earns one flattering citation in ChatGPT. Suddenly the content calendar needs to be “optimized for GEO” by Friday.

GEO is doing a lot of work for three letters.

The less exciting, and much more useful, reality is that AI-search visibility is a chain. A page must be accessible, retrieved, selected as evidence, represented accurately, cited when the interface supports citations, and connected to a business result. If you do not know which link failed, adding more content is just an expensive way to stay confused.

There is no single AI-search algorithm

Google, ChatGPT, Claude, Bing, and Perplexity do not share one index, one crawler, one ranking system, or one definition of visibility. Even within a single company, search retrieval, model training, and user-triggered browsing may use different systems and controls.

Google’s guidance for AI features in Search says its generative experiences rely on core Search ranking and quality systems. Google also describes retrieval-augmented generation and query fan-out, which can issue related searches to collect supporting information.

OpenAI documents separate crawlers for ChatGPT search, model training, and user-triggered visits. Anthropic likewise distinguishes web search from other crawler purposes. Those are not minor technical footnotes. They determine what a site can allow, block, measure, and reasonably expect.

So the first question is not, “How do we rank in AI?” It is, “Which product, which behavior, and which failure are we investigating?”

The six places AI-search visibility can break

A useful audit follows the page from publication to business outcome. Each stage asks a different question and leaves different evidence behind.

1. Eligibility: Can the system use the page at all?

Check crawler directives, server responses, firewall rules, renderability, canonical tags, indexability, and snippet eligibility. If the page is blocked or unusable, no amount of prose polishing will rescue it. The bot cannot admire writing it never received.

2. Retrieval: Does the page match the question?

Retrieval depends on whether the page is relevant to the original query and the related questions a system may use to gather evidence. Topic coverage, internal links, crawl paths, and language all matter here.

This is where traditional SEO and technical SEO still apply. They did not vanish when a chatbot learned to summarize a search result.

3. Source selection: Is the page worth using?

A system may retrieve a page and still choose another source. Original evidence, expert judgment, useful examples, corroboration, and freshness give a page a reason to win that comparison.

A generic summary has the opposite problem: it gives the system dozens of substitutes. The internet does not need another 1,800-word recap of what everyone already knows, even if the recap has excellent subheadings.

4. Synthesis: Can the information survive summarization?

Clear definitions, named entities, explicit relationships, and sensible qualifications make information easier to represent accurately. This is not permission to write for machines in tiny, joyless fragments. It is a reason to stop making readers decode vague claims.

5. Citation: Does the interface show the source?

Retrieval and citation are related, but they are not identical. A product may use information without showing a link, or cite a source without sending meaningful traffic.

Bing’s AI Performance reporting makes this distinction unusually clear: citation counts do not represent authority, ranking, importance, or placement. A citation is evidence of appearance, not a tiny trophy from the algorithm.

6. Business response: Did the visibility matter?

Track referral traffic where it exists, but do not stop there. AI answers can influence branded searches, assisted conversions, qualified visits, leads, and revenue without producing a neat last-click path.

That makes measurement harder. It does not make measurement optional.

Where marketers waste their time

They start with a universal technical hack

Google explicitly says Search does not use llms.txt, does not require special AI schema, and does not require pages to be chopped into tiny chunks. It also says there is no ideal word count for its generative Search features.

That does not prove those tactics are useless everywhere. It proves they are not universal requirements. “Add the file and hope” is not a strategy; it is a very small infrastructure project wearing a strategy costume.

They publish commodity content at scale

AI can reduce production cost. Great. That is an operating advantage, not an editorial thesis.

Before Perfect Search Media publishes an article, it should contribute something inspectable: original analysis, a decision rule, an operating method, a dataset, a qualified expert observation, or a concrete example. If the article merely restates the existing search results more efficiently, the efficiency belongs to the production team, not the reader. That standard belongs in content strategy as much as it belongs in an AI-search plan.

They treat one prompt screenshot as proof

An answer can change with the prompt, user, location, product surface, model version, and date. A favorable screenshot proves that one answer appeared once.

A screenshot is a souvenir, not a measurement system.

Prompt testing becomes useful when it is repeated, documented, and interpreted alongside crawler activity, citations, referrals, conversions, and competitive patterns.

They confuse citations with rankings

Citation volume is worth tracking. It is not a replacement for business outcomes, and it should not be dressed up as a universal rank. A page can be cited frequently and send almost no qualified traffic. Another can influence research without generating an obvious referral.

What the available research actually supports

The original Generative Engine Optimization study found that citations, quotations, and statistics improved its answer-visibility metrics by up to 40% in the study’s main benchmark. Keyword stuffing performed worse in the study’s narrower deployed-engine experiment.

That is meaningful evidence for evidence-rich writing. It is not a promise that adding one statistic will increase leads by 40%. The study measured visibility in an experimental setting, not qualified traffic, revenue, or universal performance across every production platform.

The useful takeaway is pleasantly unglamorous: support claims, make expertise visible, and avoid keyword stuffing. Marketing occasionally rediscovers editing and gives it a new acronym.

A better way to allocate AI-search work

Perfect Search Media’s recommended sequence starts with the failed stage, not the fashionable tactic:

• Confirm access and eligibility. Review crawler controls, server behavior, rendering, canonicalization, indexability, and platform-specific requirements.
• Define the decision and query set. Identify the questions that matter to the intended buyer and the related questions a retrieval system may explore.
• Evaluate the available sources. Look for original facts, expert judgment, useful examples, and current information competitors do not provide.
• Test retrieval and citation repeatedly. Record the platform, prompt, date, cited sources, answer pattern, and meaningful variation.
• Connect visibility to outcomes. Review referrals, branded demand, assisted conversions, qualified visits, leads, and revenue.

Use the Perfect Search Media AI Search Failure Map

The failure map turns a vague visibility concern into a next action. If eligibility fails, fix technical access. If retrieval fails, revisit relevance, coverage, and authority. If source selection fails, strengthen the evidence and distinct value. If synthesis fails, clarify the entities, relationships, and limitations. If citation fails, compare source patterns without pretending citation is a rank. If business response fails, review the query value, landing experience, offer, and measurement.

It also creates a useful stopping rule: if the evidence does not identify a failure, do not prescribe a cure. Sometimes the next project is better measurement, not another article.

What marketers should do next

Choose one commercially meaningful query group and trace it through all six stages. Document what is accessible, what is retrieved, which sources appear, what the answer says, whether a citation is visible, and what happens after exposure.

That review will produce a narrower, more defensible plan than a list of universal GEO best practices. It may also reveal that the problem is technical, editorial, analytical, or simply not proven yet. All four answers are more valuable than guessing.

Perfect Search Media applies this diagnostic across AI search, content strategy, and technical SEO. The goal is not to promise control over an answer engine. It is to identify what a brand can improve, separate it from what nobody can verify, and invest accordingly.

Need a clearer diagnosis before your next AI-search investment? Explore Perfect Search Media’s AI Search services.

Frequently asked questions

Is there one algorithm for AI search visibility?

No. Google, ChatGPT, Claude, Bing, and Perplexity do not share one index, crawler, ranking system, or definition of visibility. Investigate a specific product and a specific failure, not a universal rank.

Do llms.txt, special schema, or short content chunks guarantee citations?

Google says its generative Search features do not use llms.txt, do not require special AI schema, and do not require pages to be split into tiny chunks. Those tactics are not universal requirements, and none of them guarantees retrieval, citation, or traffic.

Does a citation count as a ranking?

No. A citation shows that a source appeared in an answer interface that supports citations. Microsoft’s Bing AI Performance reporting states that citation counts do not represent authority, ranking, importance, or placement.

Sources

• Google Search Central, “Succeeding in Google’s AI search experiences”:https://developers.google.com/search/docs/fundamentals/ai-optimization-guide
• OpenAI, “Overview of OpenAI Crawlers”:https://developers.openai.com/api/docs/bots
• Anthropic, “Web search API”:https://claude.com/resources/articles/web-search-api
• Microsoft Bing Webmaster Blog, “Introducing AI Performance in Bing Webmaster Tools”:https://blogs.bing.com/webmaster/2026/2/Introducing-AI-Performance-in-Bing-Webmaster-Tools-Public-Preview/
• Aggarwal et al., “GEO: Generative Engine Optimization” (arXiv:2311.09735):https://arxiv.org/abs/2311.09735