Free lesson

The New Reality: How AI Chooses What to Recommend

The New Reality: How AI Chooses What to Recommend — LLM SEO: Get Your Product Discovered in AI Answers. From a list of links to a composed answer; Why clicks no longer tell the whole story; What a model knows and what it retrieves now; What may influence a mention; New measurements without abandoning SEO; Build a baseline you can compare. Practical…

Demo edition. The materials are undergoing editorial review.

Topics covered

  • From a list of links to a composed answer
  • Why clicks no longer tell the whole story
  • What a model knows and what it retrieves now
  • What may influence a mention
  • New measurements without abandoning SEO
  • Build a baseline you can compare

From a list of links to a composed answer

In conventional search, a person receives a results page, opens several links, and builds a shortlist. An AI assistant can perform part of that work: answer a full question, name several products, and explain where they fit. That does not make every answer a final or correct choice. The output can depend on the wording, access to web search, retrieved sources, and conversation context.

AEO, or Answer Engine Optimization, means working so that a product is represented accurately in a composed answer and, when the system displays sources, supported by a relevant citation. It complements SEO rather than replacing it: Google explicitly says that fundamental SEO practices remain relevant to AI Overviews and AI Mode.

Fictional example. Ledgerly is a bookkeeping app for freelancers who invoice in several currencies. It is absent from an answer to “What is the best bookkeeping app for a freelancer?” That does not yet prove a product problem. First, test a more precise use case, several systems, and the sources behind their answers.

Why clicks no longer tell the whole story

A person may see a product in an AI answer, avoid clicking, and search for the brand or type its address several days later. Analytics may then record branded search or direct traffic. That journey is possible, but without supporting evidence you cannot attribute every increase in those visits to AI.

Use several signals together: referral visits from assistants, answers to “How did you hear about us?”, changes in branded search, and actual mentions in saved AI answers. Even a direct referral does not reveal the whole journey, while a survey answer depends on the user’s memory.

Common mistake: counting only visits with an assistant’s domain as referrer. The opposite mistake is assigning all direct traffic to AI. Both replace measurement with an assumption. Keep the signals separate and look for a consistent change without presenting correlation as proven acquisition.

What a model knows and what it retrieves now

It helps to separate two situations. Without web search, a language model produces an answer from patterns learned during training and the current context. That knowledge may be incomplete or outdated. With search enabled, a system can formulate search queries, review current pages, and attach links. OpenAI explicitly warns that search citations can still be incomplete, outdated, or incorrect, so readers should open and verify them.

Finding material and using it in a response is often called retrieval; the broader architecture is commonly called RAG, or retrieval-augmented generation. Not every recommendation uses the same mechanism, and a visible citation does not explain the entire ranking process.

For Ledgerly, this creates two different questions: can the model describe the product accurately at all, and does search retrieve a current page about multi-currency invoicing? One new post cannot instantly rewrite learned model knowledge; the retrieval question can be investigated through page accessibility, content, and displayed sources.

What may influence a mention

There is no public universal formula that guarantees a recommendation across assistants. A practical diagnosis begins with four questions. Relevance: does the material name the intended audience and job? Verifiability: are precise facts available on pages a reader can open? Independent support: is the product discussed on relevant third-party sources rather than only its own site? Freshness: do those materials describe the current product?

These are diagnostic considerations, not published algorithm weights. A company page matters as a primary source for features, prices, and limitations. An independent review or substantial discussion can add comparative context, but a random comment does not become authoritative merely because it is external.

For Ledgerly, “simple accounting for everyone” supplies little context. “Invoicing and bookkeeping for freelancers paid in EUR and USD” matches a specific need more closely. It is useful only if the product actually supports that workflow and the page states its conditions and limits.

New measurements without abandoning SEO

Rankings and traffic still matter: a page must be found, opened, and understood. Add measurements of AI answers. Mention share is the number of saved answers naming the brand divided by all tested answers. Mention position records where it appeared in a list, but it is not a stable “ranking.” Context records the use case and whether the stated facts are accurate. Citation separates a brand name from a visible source link.

Fictional calculation: test 10 questions in three systems, producing 30 answers. Ledgerly appears in 6. Its mention share in this specific test is 6 ÷ 30 = 20%. This is not market share; it describes only the chosen prompts, systems, date, and test conditions.

Keep platform results separate. Practitioner discussions on Reddit similarly recommend saving full answers, mentions, and cited URLs because outputs can change. That is a useful field method, not an official measurement standard.

Build a baseline you can compare

A baseline is a dated record of how selected systems answered a fixed set of buyer questions. Write 10 prompts covering the category, audience, specific job, competitor alternatives, comparison, price, and constraints. Run each in a new conversation in ChatGPT, Perplexity, and Gemini. Record the date, system, exact prompt, complete answer, brand presence, order of names, stated use case, factual errors, and displayed sources.

Do not merge mentions and citations: “Ledgerly suits freelancers” without a link is a mention; an adjacent clickable source is a citation. Do not infer a pattern from one run or compare results after silently changing prompts. Personalization, region, search mode, and generation variability can affect an answer.

The initial snapshot does not prove that an optimization worked. It creates a reference point. Later, repeat the same conditions, disclose any method changes, and compare the accuracy of the product description as well as the number of mentions.

Practical assignment

Save your AI-answer visibility baseline

Continue to learning

Course outline

  1. The New Reality: How AI Chooses What to Recommend
  2. The AEO Framework: OCEAN
  3. What ChatGPT, Perplexity and Gemini “See” About Your Product
  4. Content Strategy for AI Discovery
  5. Technical Optimizations: Crawlers, Structured Data, llms.txt
  6. Case Study: How Tally Made AI Its Top Acquisition Source
  7. Case Study: How Vercel Adapted SEO for LLMs
  8. Brand Mentions and Citability
  9. Monitoring: How to Track AI Recommendations
  10. AI SEO + Classic SEO Synergy
  11. Common Mistakes and Myths of LLM SEO
  12. Your 30-Day LLM SEO Action Plan