How to Rank in AI Search

Every guide on this topic tells you to rank in AI search, and the word rank is what makes them useless. We run outbound for 50+ B2B companies and have sent over 8 million personalized emails this year, so we watch how buyers behave before they reply, and a growing share of them ask an assistant about you first. Below is what the engines are actually doing, the source pool each one draws from, the 4 levers that decide whether you get named, and the measurement routine that tells you whether any of it moved.

What does it mean to rank in AI search?

Search trained everyone to think in slots. You were 4th, you worked, you became 2nd, and the click volume moved in a predictable way. That mental model does not survive contact with a generated answer.

An assistant produces one paragraph. Either your company is inside it or it is not. Reorder the same run and the names inside it shuffle anyway, because generation is probabilistic rather than fixed. So position is not a real measurement, and presence is.

The second thing that breaks is the idea of a page competing for a keyword. The engine is not choosing between your page and a competitor page. It is assembling a claim about a category and deciding which companies belong in it, then attaching a few links to support the sentence. You are not competing for a slot on your own site. You are competing for a clause in somebody else's sentence.

AI search ranking

A category error. There is no ordered result set inside a generated answer, so there is no rank. What people mean by the phrase is presence: how reliably an assistant names a company when asked a question its buyers would ask.

Source pool

The set of domains an engine leans on when it builds an answer in a given category. Each engine has a different one, which is why a single content plan produces uneven results across ChatGPT, Perplexity, and Google AI Overviews.

If you want the full definition of the discipline, that sits in what generative engine optimization actually is, and the differences from the old job are worked through in GEO versus SEO.

Why is there no ranking to win?

Because the retrieval step and the writing step are separate, and only the first one looks anything like search.

When a buyer asks an assistant for the best outbound agencies for consultancies, the engine pulls a set of documents, then writes a paragraph from what it already believes plus what it just read. The paragraph is not a ranked list of the documents. It is a summary of a consensus, and a company gets into that consensus by being described the same way in enough places that the description reads as settled fact.

This is why the correlation data keeps pointing away from links. Ahrefs analyzed 75,000 brands and found brand web mentions correlated with AI Overview visibility at 0.664, against 0.218 for backlinks. A mention with no link at all can move your presence, and a link from a page nobody describes you on may move nothing.

It is also why the answers are unstable in a way search results never were. A study in the Journal of General Internal Medicine ran identical prompts through 6 commercial models 5 times each and documented real variation in what came back. One screenshot proves nothing. Twenty runs of the same prompt is a number.

You are not competing for a slot on your own website. You are competing for a clause in somebody else's sentence.

How does an AI engine actually pick its sources?

In 4 steps, and only the second one resembles search. Understanding the order is what stops you from spending a quarter on the step that decides the least.

1. It rewrites your question into several. One buyer question becomes a handful of narrower queries the engine runs behind the scenes. Ask for the best podcast agency for a consulting firm and the engine goes looking for agency lists, pricing pages, reviews, and category definitions separately. You are not matched against one phrase, you are matched against a spread of them.
2. It retrieves documents. This is the only step that behaves like an index, and for the engines that browse live, standard crawlability is the whole entry fee. Google is unusually direct about this in its own documentation on AI features: there are no additional requirements to appear in AI Overviews or AI Mode, and no special schema.org structured data you need to add.
3. It grounds the answer in passages, not pages. The engine pulls blocks out of what it retrieved, so a page is never cited whole. A self contained paragraph gets lifted. A paragraph that only makes sense after the 3 above it gets skipped in favor of a competitor's cleaner one. That is the mechanic behind answer capsules, comparison tables, and the content assistants skip.
4. It writes from what it already believed, plus what it just read. The model arrives at the question with an existing impression of your category and the companies in it. Retrieval nudges that impression. It does not replace it. Which is why a brand nobody discusses anywhere can publish a perfect page and still not get named.

Step 4 is the one that reorders everyone's priorities once you take it seriously. The model's prior comes from the open web describing you, and that is built outside your domain over quarters. Steps 2 and 3 are your site, they are cheap, and they only decide how you get quoted once the model already thinks you belong.

Retrieval working at the passage level also changes how you structure a site rather than a page. Blocks that stand alone, headings that match the question, and internal linking built for retrieval rather than for crawl budget. The depth of coverage across a whole subject is topical authority for AI search, and the maintenance question is how often to refresh content.

Which sources does each AI engine actually pull from?

This is the part almost every checklist skips, and it is the difference between a plan that works on one surface and a plan that works on 3.

Profound studied 30 million citations across ChatGPT, Google AI Overviews, and Perplexity between August 2024 and June 2025. The headline finding is that the 3 engines do not agree on who to trust. Wikipedia was the single most cited domain for ChatGPT at 7.8% of its citations. Reddit led for both Perplexity at 6.6% and Google AI Overviews at 2.2%. Search Engine Roundtable covered the split when it published.

Read that as a routing problem rather than a ranking problem. The same buyer question hits 3 engines that reach for 3 different kinds of evidence, so the work splits accordingly.

The pattern underneath the table matters more than the individual rows, because the percentages will move. Every engine is reaching for evidence produced by somebody other than you. Your own site is the reference the engine checks after it has already decided you belong in the answer.

What actually gets a company named?

Four levers, in the order they pay.

1. Third party descriptions. The number of independent domains that say what you do, in language that matches how buyers ask. This is the lever with the strongest correlation and the slowest payback, which is why most teams skip it for the fast ones.
2. Description consistency. The same category words in every place you appear. If your site says invitation led outbound, your LinkedIn says lead generation, and a directory says marketing agency, you have handed the model 3 conflicting labels and it will pick the vaguest one.
3. Original data. A number nobody else can source makes you the only citable answer to a question. There is no way to be the source of a benchmark you did not measure, which is exactly why our cold email reply rate benchmarks, podcast lead generation benchmarks, and state of AI outbound exist.
4. Extractable structure. Self contained blocks, a direct answer in the first 120 words, plain schema, and headings phrased as the question a buyer would type. This does not get you onto the shortlist. It decides whether you get quoted once you are already on it.

Lever 3 has the only controlled evidence in this whole discipline behind it. The Princeton team that named the field tested content changes against 10,000 real queries across 9 domains in GEO: Generative Engine Optimization and found that adding statistics and quotations from credible sources lifted visibility in generated answers by up to 40%, while keyword stuffing moved nothing. Specific sourced numbers are what an assistant reaches for, because a number is the part of a sentence it cannot paraphrase from somewhere else.

Which gives you a sharper version of the same lever. A stat you borrowed makes the original publisher the citable source. A stat you measured makes you one. That is the only reason to run your own benchmark study instead of quoting somebody else's.

Order matters because the effort is inverted. Structure takes an afternoon and moves the least. Third party descriptions take quarters and move the most. Most companies do the afternoon, watch nothing happen, and conclude the whole channel is hype.

The mechanics of each engine's shortlist decision are broken down in why AI answers cite some companies and not others, and the surface specific tactics in getting cited by ChatGPT and Perplexity, getting cited by Perplexity, optimizing content for ChatGPT, and AI Overviews for B2B. Keeping one description across every surface is entity consistency.

How do you earn mentions on domains you do not own?

By producing artifacts other people publish. That is the whole job, and there are only a few honest ways to do it at any volume.

Recorded interviews are the highest ratio version. One conversation produces a transcript page, a video with a title and description, an episode page on the host's site, clips, and usually a post from each side. That is 5 or 6 independent artifacts describing your company in your own words, on domains you do not control, out of a single hour. The retrieval mechanics are in podcast transcripts for AI search and how to get your podcast cited by AI.

Running the show yourself flips the direction of the work. Instead of pitching to be a guest and waiting on somebody else's calendar, you invite the people you want to be associated with, and every recording produces the same artifact stack with your name attached to all of it. The mechanism is podcast led outbound, the operating version is a podcast acquisition system, and the reason senior people accept is in why executives say yes to podcast invites.

The unglamorous layer underneath decides whether any of it happens. An invitation only reaches a senior buyer if the sending setup clears filters, which is email deliverability and ongoing deliverability monitoring. Who you point it at is the ideal customer profile question, worked through for this motion in building the guest list. And the multiplication pass that turns one recording into a dozen artifacts is in repurposing episodes.

What should you actually do on your own site?

Less than the checklists claim, and it is not nothing. The honest framing is that your site cannot get you onto the shortlist, and it can absolutely keep you off it.

Start with crawl access, because it is binary. Two of the engines in the table above retrieve live pages, so being fetchable is a prerequisite rather than a tactic. GPTBot is documented in OpenAI's crawler reference, Anthropic publishes ClaudeBot and how to block it, and Perplexity lists its agents in its bots guide. Blocking them is a reasonable call for a publisher protecting paid archives and the wrong call for a B2B company that wants to be named in a buying conversation. The full breakdown is in robots.txt for AI crawlers, and it is an afternoon of work, so do it before you spend a quarter on anything else.

Everything after that is a question of how much credit the layer deserves. Here is the honest version, lever by lever.

Write for extraction and the first 3 rows take care of themselves. One idea per block, the answer before the setup, and headings phrased the way a buyer types. If a paragraph only makes sense after reading the 3 paragraphs above it, an assistant cannot lift it, so it will lift a competitor's cleaner version instead. The per element versions are in FAQ schema for AI visibility, schema markup for podcast pages, and building an llm-info page.

Be careful about how much credit you give the whole layer. A Stanford audit of generative search engines, Evaluating Verifiability in Generative Search Engines, found only 51.5% of generated sentences were fully supported by the citations attached to them. Being linked is not the same as being described correctly, which is why the audit routine in auditing your brand in ChatGPT tracks the wording and not just the mention. As for llms.txt, the spec lives at llmstxt.org as a community proposal rather than an adopted standard. Publish one if you like, and do not report on it as a metric.

How do you measure citation share?

By running the same questions over and over and counting. There is no console for this, and the absence of one is why so many teams report on the wrong number.

Citation share

The percentage of runs of a given buyer question where an engine names your company. If a category prompt names you in 6 of 20 runs, your citation share on that prompt is 30%. It is the only metric in AI search that behaves like a measurement rather than an anecdote.

Prompt panel

A fixed set of 15 to 25 questions your buyers would actually type, written once and never edited, so every later reading is comparable to the baseline. Changing the panel resets the measurement.

The routine is 4 rules, and all 4 exist to kill a specific way of fooling yourself.

1. Fix the panel before you measure anything. Write the 15 to 25 questions, save them, and stop touching them. A panel you tune as you go always trends up.
2. Run every prompt 3 to 5 times per engine. Generated answers vary run to run. A study in the Journal of General Internal Medicine put identical prompts through 6 commercial models 5 times each and documented real variation in what came back. One screenshot is not a reading.
3. Use clean sessions with memory turned off. An account that has been chatting about your company for a year will name your company. That tells you nothing about a buyer who has never heard of you.
4. Log 3 things, not 1. Whether you were named, the exact words used to describe you, and which competitors appeared alongside you. The description is the field that changes first, and it is the one that predicts the rest.

Do not substitute referral traffic for citation share. The engines are built to answer rather than to forward, and Pew Research measured the gap across 68,879 Google searches from 900 US adults: users who saw an AI summary clicked a traditional result in 8% of visits, against 15% when no summary appeared. Being the named answer and getting the click are now two different outcomes, and only one of them shows up in analytics. The full method is in how to track AI search visibility, and the tradeoff is argued out in LLM citation versus SEO traffic.

How long does it take?

Quarters. Record a baseline first or you will have no way to tell the difference between progress and model noise.

Expect flat readings for the first 2 months, because the mentions have to be published, crawled, and repeated before they shift anything. Real movement looks like a category prompt that named you in 3 of 20 runs in January naming you in 8 of 20 by June, or a description that read as a generic tool in month one reading as your actual category by month five.

Judge the traffic on quality rather than volume when it does arrive. Semrush found in its AI search traffic study that the average visit from an AI source converted at roughly 4.4 times the rate of the average visit from traditional organic search, and its 17 month clickstream analysis put outbound referral traffic growth from ChatGPT at 206% in 2025. Small numbers, late stage intent, because the assistant already did the shortlisting before the click. Treat that traffic as the receipt rather than the goal, and keep the citation share panel as the scoreboard.

One more reading worth taking at month 6. Count how many of the pages an engine cited when it named you sit on domains you do not own. If the answer is all of them, the engine is describing you from other people's work and the plan is working the way the evidence says it should. If the answer is none of them, you are being quoted by the only publisher who talks about you, which is yourself, and that is a position one competitor press mention can take from you.

Where this lands

There is no rank to win, so stop building a plan around one. There is a description of your company circulating on domains you do not control, and the engines are summarizing it. Your job is to make that description accurate, specific, and common enough that summarizing it produces your name.

The uncomfortable part is that this is a publishing and relationship problem wearing a technical costume. The teams winning it are not running better schema. They are appearing in more conversations, saying the same thing every time, and publishing numbers nobody else has.

Recorded conversations happen to be the cheapest way to produce all 3 of those at once, which is the long way round to what a podcast does for your search footprint and turning episodes into answer content.

Which is also why we do not sell AI visibility as a service. We run an invitation engine that puts our clients in recorded conversations with their ideal buyers, and it carries one commitment: 30 recorded conversations in 90 days, or your money back. Editing is included, the recordings are theirs, and the invitations go out over email only. The AI visibility is what falls out of running it, not the reason to run it.

Set the baseline panel this week, fix your robots.txt this afternoon, and put the rest of the quarter into being described by other people. That is the whole strategy, and the only part of it anyone can copy is the afternoon.