The firms getting quoted by ChatGPT and Perplexity are not the ones publishing the most. They are the ones a language model can read without guessing.
Most teams met AI search the way they meet every shift in the industry: make more content. Another explainer. Another 2,000-word guide. Another listicle nobody asked for. It moved nothing, because volume was never the constraint.
How to get cited by AI search is a data problem, not a blogging problem. A model names a source when it can extract a clean, self-contained, attributable answer from the page. When your sharpest expertise is buried in prose a parser has to interpret, it goes uncited while a thinner competitor with cleaner structure gets the mention.
The objection comes fast: surely a genuinely useful article still wins on merit. It does not, and the reason is mechanical. Merit the model cannot extract is merit it cannot cite. A brilliant argument spread across nine dependent paragraphs is invisible to a system that quotes in passages, while a mediocre page that answers the exact question in one clean line gets lifted whole.
AI Search Retrieves, It Does Not Rank
Google spent twenty years teaching marketers to think in rankings. Ten blue links, position one, climb the list. AI search does not work that way, and treating it like Google is the first mistake firms make.
When someone asks ChatGPT or Perplexity a question, the system does not hand back a ranked page of options. It retrieves passages from many sources, represents them as vectors, and synthesizes one answer. Your page is competing to be a passage worth lifting, not a link worth clicking.
Here is the mechanism most teams never picture. Your page gets split into chunks, each chunk turned into a vector, each vector stored so it can be matched against a question. The model pulls the chunks closest in meaning to the query and builds its answer from those. A page written as one long argument with no clean internal boundaries chunks badly, and a badly chunked page rarely surfaces.
That changes the unit of value. It is no longer the article. It is the paragraph, the definition, the table row, the clearly labeled fact. If a model cannot isolate a clean chunk of meaning from your page, your page does not exist to it, no matter how well it ranks in classic search.
Make one move this week to feel the difference. Take your best page and ask whether any single paragraph, read cold with nothing above it, answers a real buyer question on its own. If it does not, you have written for a reader who scrolls, not a system that retrieves, and the two now want opposite things from the same page.
How to Get Cited by AI Search Starts With Schema, Not Copy
Structured data is how you tell a machine what your words mean instead of forcing it to infer. It is JSON-LD in the head of your page, and it is the single highest-leverage move most firms have not made.
Prose is ambiguous to a parser. A sentence that says a firm has represented clients across the Gulf South since the early nineties is clear to you and a guess to a model. The same fact expressed as Organization schema, with a foundingDate and an areaServed, is a fact the system can read, store, and repeat with confidence.
You will hear that schema is overrated, that Google has said structured data is not a ranking factor, that it changes nothing. That advice is about classic search, and it is aimed at the wrong machine. AI retrieval does not need schema to rank you; it uses schema to trust you. The question is not whether markup lifts a position. It is whether a model can attribute a claim to you without hallucinating, and structured data is the difference between a fact it will repeat and a guess it will avoid.
Confidence is the whole game. Models cite what they can attribute without risk. Structured data lowers the risk of quoting you, so you get quoted.
Do it on the pages that carry your real expertise first. The service pages, the deep explainers, the pages a buyer reads before a call. Those are the pages a model reaches for when someone asks who does this work well, and those are the pages most firms leave completely unstructured.
The Markup That Actually Earns Citations
Not all schema is equal. A few types do most of the work, and they map directly to the questions AI search fields all day.
- Organization and Person markup that names who you are, links to your profiles with sameAs, and ties every article to a real author. This is how a model learns you are an entity, not a URL.
- FAQPage schema that pairs an explicit question with a self-contained answer. This is the most liftable format that exists, because it hands the model the exact shape it wants to output.
- Article schema with a headline, datePublished, and a dateModified you actually keep current. Freshness is a citation signal, and a stale date reads as a stale source.
- HowTo schema for anything procedural, so a step-by-step becomes a structured sequence rather than a wall of paragraphs a parser has to segment.
- BreadcrumbList and clear canonical tags, so the system knows which URL is authoritative and never splits your credibility across duplicates.
The mechanism behind FAQPage is worth sitting with, because it is the clearest case. When you mark a question and its answer as a pair, you are not decorating the page. You are handing the model a pre-built question-and-answer unit in the exact format its output takes. It does not have to segment your prose, guess where the answer starts, or decide how much context to keep. You have done the chunking for it, and a chunk you shaped is a chunk you control.
None of this is exotic. It is a few blocks of JSON-LD and a discipline about keeping it accurate. The firms that treat it as engineering, not marketing, are the ones getting named.
Publishing more articles into a site a model cannot parse is like shouting louder in a language nobody in the room speaks.
Write Passages a Model Can Lift Whole
Markup gets you read. Structure inside the copy gets you quoted. The two work together, and most sites fail the second one.
Write answer-first. Put the claim in the opening sentence of a section, then support it. A model scanning for a passage to quote wants a sentence that stands on its own without the three paragraphs above it for context.
Shape your H2s as the questions people actually ask. Define your terms in single, clean sentences a parser can extract as a definition. Break procedures into lists. Every one of these choices makes a passage more liftable, and every buried claim makes you easier to skip.
Here is the test for a liftable passage. Copy one sentence out of a section, paste it into a blank document, and read it as a stranger would. If it still states a complete, attributable fact, a model can quote it. If it collapses into 'this' and 'that' and 'as mentioned above,' it was written to be read in sequence, and sequence is exactly what retrieval strips away.
This is the same lesson the piece on this blog arguing your ad is not underperforming, your creative is boring drove home about paid media. The channel is rarely the problem. The asset you put into it is. Undifferentiated copy loses in a paid auction and it loses in a model's retrieval step for the same reason: nothing about it demands to be chosen.
How to Get Cited by AI Search Without Domain Authority
Small firms assume citations go to the biggest domains. They go to the clearest entities. A regional firm with tight, consistent structured data can outcite a national brand whose markup is a mess.
The instinct to give up here is strong. A national competitor has ten thousand backlinks and a decade of press, so why fight a machine that surely rewards size. Because the machine is not counting your links, it is checking whether your identity holds together. A firm whose name, leadership, and location are described one way on the site and three other ways across the web hands the model a reason to doubt, and doubt is what keeps you out of an answer.
Entity consistency is the mechanism. Your name, your organization, and your people should be described identically everywhere a model can find them: your site, your profiles, your directory listings, your press. When a model sees the same entity corroborated across independent sources, it trusts the attribution and cites it.
The sameAs property is the underused lever here. It links your Person and Organization entities to your other verified presences, and every link is a thread the model follows to confirm you are who you say. A firm with five corroborating links is easier to cite than one with none, because the model can check its own work before it names you.
This is exactly why the post introducing the team behind the productions matters more than it looks. A model wants to bind a claim to a real, verifiable person with a role and a history it can cross-check. Named authors with Person schema and linked profiles turn anonymous content into attributable expertise, and anonymous content is the easiest thing for a model to leave out.
Measure Citations, Not Clicks
The last reason firms lose at AI search is that they never notice they are losing. Their dashboard tracks organic sessions, and AI answers often produce zero clicks while still shaping the buyer's shortlist.
Change what you watch. Query the major models directly for the questions your buyers ask and record whether you appear, how you are described, and who gets named instead. Track share of model answers the way you once tracked share of voice.
- Ask ChatGPT, Perplexity, and Google's AI answers your top ten buyer questions on a schedule and log who gets cited.
- Check whether the description the model gives of you matches the facts in your schema, and fix the schema when it does not.
- Watch dateModified across your key pages, because a source that never updates slides out of the answer set.
Turn this into a standing ritual, not a one-time audit. Pick the ten questions a buyer types before they call you, run them through every model on the first Monday of the month, and paste the answers into a sheet beside last month's. The trend line is your real scoreboard. Watching your name appear, then get described more accurately, then start beating a competitor you used to trail tells you the plumbing is working long before any click does.
You cannot fix what you refuse to measure. A firm that only counts clicks will keep concluding AI search does not matter, right up until it has been quietly written out of every answer in its market.
What Changes Monday
Stop the content calendar for a week. The next article will not get you cited if the plumbing underneath it cannot be read.
Open your highest-value page, the one that describes what you actually sell, and audit it. Does it carry Organization and Person schema. Does the author resolve to a real profile. Does the top of each section answer a question in one clean sentence. Is the dateModified honest. Most firms fail four of those five checks on their most important page.
Then fix them in order: schema first, author entities second, answer-first passages third, measurement last. Add FAQPage markup to the pages that already field real questions. Point every author to a consistent, corroborated identity. Rewrite the buried claims into sentences a model can lift whole.
The work is technical, repeatable, and productizable, which is why AI search visibility belongs alongside media strategy and podcast authority as a service worth buying rather than a mystery worth guessing at. Getting cited by AI search is not about writing more. It is about making what you already know machine-readable, attributable, and impossible to leave out.