The AEO playbook: how to get cited by ChatGPT and Perplexity

The AI search traffic shift is not the one most teams are preparing for. The common framing (ChatGPT replaces Google, so marketing moves from SEO to AEO) misses two findings in the hard data from the last twelve months. First, AI referral is diversifying fast: ChatGPT's share of measurable AI referrals fell from 89.1 percent to 62.6 percent between the May to August 2025 window and the March to April 2026 window, with Claude moving from 1.4 to 18.5 percent, Gemini from 2.4 to 10.6 percent, and Perplexity from 3.1 to 7.3 percent, per Higoodie's AI Search Traffic Report 2026 based on a panel of 41 mostly B2B brands. A team optimizing for a single AI platform is already a step behind the mix.

Second, visibility concentrates. Semrush's AI Visibility Index 2026 analyzed 126 million US AI search prompts across 22 industry verticals and more than 1,200 brands between January and April 2026, and found that only 36 brands appeared in the top 100 most-mentioned list on every tracked platform in every month. That is roughly 3 percent of the brands studied. Average citations per response differ sharply by platform: 15.4 on ChatGPT, 11.4 on Google AI Mode, 9.2 on Google AI Overviews, 3.3 on Gemini. Which platform the buyer uses decides how many sources the answer draws from, and whether yours is one of them.

This post is for the US B2B marketing leader deciding what to fund next quarter and for the developer who will implement whatever that decision is. The sections below name the retrieval mechanism AI engines use to pick what to cite, the six AEO levers that measurably move citation share, the schema markup that AI engines actually read (and the schema types to skip), how to measure AEO in a world where Google rankings are no longer the signal, what AEO does not fix, and a 60-day plan that stays practical. Where the implementation is deeply technical, the next sentence translates for the marketing leader reading along.

AI search is not a search engine with better formatting. It is a retrieval-augmented generation system over the web. Each platform maintains an index of pages it can see, a retrieval model that fetches the most relevant chunks for the query, and a large language model that composes an answer and cites the sources it leaned on. The ranking paradigm from classic SEO, where one query produces one ranked list of blue links a user clicks, does not describe how any of this works.

Three stages decide whether your page gets cited. First, retrieval: the platform's crawler has to have seen your page, the content has to be extractable from the HTML as served (not injected after render), and the chunks have to be semantically findable for the query. Second, relevance: the retrieval model picks a small set of chunks, typically five to twenty, to pass to the LLM. Third, synthesis and citation: the LLM composes the answer and attributes claims to specific retrieved sources, with citation counts varying by platform as above.

Each platform runs its own index and retrieval. Google AI Overviews and AI Mode draw from the Google search index: Google's search team confirmed on May 15, 2026 that AI features draw from the same index as classic search, which means a page not findable in Google will not surface in AI Overviews. ChatGPT draws from a combination of Bing's index, OpenAI's own crawl (GPTBot and OAI-SearchBot), and specific publisher partnerships. Perplexity runs its own index. Claude uses Brave Search and other partners. The practical consequence: a page that is crawlable, extractable, and ranks reasonably in classic search is the baseline for appearing anywhere; AEO is the discipline that moves it from retrievable to cited.

In plain English for the marketing leader: the LLM is not browsing your site, deciding how impressive the brand looks, and citing you because the brand is strong. It is reading a short bundle of extracted chunks that its retrieval model picked, and quoting the ones that most directly answered the question. The failure modes here are the same ones that break a [RAG system in production](/blog/why-rag-fails-in-production): a related chunk retrieved instead of the right one, a confident answer composed from the wrong source, a chunk that was never retrievable in the first place because the content was injected after render. The work of AEO is to make the chunks that get retrieved be yours, and to make them the ones that most directly answer the question.

Six levers measurably move the probability that an AI engine cites your page. None of them is new to classic SEO. All of them matter more in AI search because the LLM is picking a tiny slice of your page to quote, and the slice has to be good.

First, direct-answer architecture. The one-sentence answer to the query goes in the first paragraph, in declarative form, with the key noun phrase in it. The retrieval model is more likely to pick a chunk that sits at the top of the page, and the LLM is more likely to quote a sentence that stands alone. A page that opens with context, caveats, or an anecdote has already lost half the citation chance. The marketing version of this: write a sentence that would survive being pulled out of context without a reader needing to see the headline.

Second, entity consistency. Your Organization schema, your canonical URL, your logo file, your name across LinkedIn, Crunchbase, Wikipedia (where applicable), and your authoritative mentions on third-party sites all have to tell the same story. AI engines maintain an entity graph that disambiguates brands with similar names, and inconsistent entity signals are the most common reason a brand gets cited less than its content quality predicts. The Organization JSON-LD block (name, url, logo, sameAs linking to LinkedIn, GitHub, press mentions) is the structured signal that anchors the graph.

Third, Article plus FAQPage plus BreadcrumbList JSON-LD. These three schema types do most of the AI-citation work on a content page. Article (or BlogPosting) names the page as content with an author, a publication date, and a publisher. FAQPage emits the question-and-answer pairs as structured data, which both Google and OpenAI documentation list as a format AI retrieval models privilege. BreadcrumbList gives the LLM a navigational hierarchy to describe where this page sits in the site. Rich snippet types that Google has deprecated or that AI engines ignore (standalone HowTo for non-recipe content, Review and Rating schema without real reviewed products, QAPage for non-forum pages) add noise without upside.

Fourth, author and publisher authority signals. For B2B content, the Article's author field should resolve to a Person schema object with a URL, a sameAs array, and credentials; the publisher field should resolve to the Organization. For sites that publish as the organization rather than individual authors (Stallwart's own model is one example), both author and publisher resolve to the same Organization, and the entity consistency work above does the authority lifting. Fabricated author bios or invented credentials are a documented failure mode; AI engines cross-reference against the public web, and an author with no corroborating footprint signals untrusted content.

Fifth, FAQ blocks. Explicitly answering the three to eight questions a buyer would ask on this topic, each as a self-contained block with a question headline and a two-to-four-sentence answer, is the highest-leverage structural pattern for ChatGPT and Perplexity citations. The lead sentence of each block is designed to be quotable in isolation. The retrieval model picks chunks that look like answers to a question, and the clearest form of a question-answer pair is one explicitly structured as such.

Sixth, crawler-friendly rendering. The content that gets retrieved is the content in the HTML as served, not the content that renders after a client-side JavaScript run. A single-page app that renders its article body in the browser is a page that AI crawlers often see as empty. The fix is server-side rendering or static generation of the article body, with the structured data in the initial HTML. OpenAI's GPTBot, Anthropic's ClaudeBot, Perplexity-User, and Google's AI-crawler variants all respect robots.txt; a disallow rule accidentally blocking them means the content is invisible to AI search by configuration.

In plain English for the technically literate operator: these six levers are not a mysterious new discipline. They are the specific moves that make a page legible to a retrieval-augmented LLM, and most teams have been shipping one or two of them without the others. The gains come from doing all six on the same page.

Schema markup is where most AEO initiatives either overreach or underdeliver. The useful set is small; the ignored set is large. The four JSON-LD blocks that move citation outcomes are Organization at the site level, Article (or BlogPosting) per content piece, FAQPage per page with question-and-answer content, and BreadcrumbList for navigational context. Everything beyond this list is a case-by-case decision, not a default.

Organization is site-wide. The required fields are name, url, and logo; sameAs is where the entity-graph work pays off, with URLs linking to your LinkedIn company page, Crunchbase profile, authoritative press mentions, and official social accounts. The schema is served from every page of the site. Teams running multiple brands under one legal entity should match the publishing brand's schema to the content's audience, not to the parent company.

A minimal and sufficient per-page JSON-LD block uses Article (or BlogPosting; both validate) as the type, with the fields headline, description, datePublished, dateModified, mainEntityOfPage, author, publisher, image, and articleSection. The author and publisher resolve to typed objects (Person or Organization) with their own url and name. BreadcrumbList lives alongside Article and uses an itemListElement array with position-numbered entries pointing at each level of the navigational hierarchy. FAQPage uses mainEntity as an array of Question objects, each with its own acceptedAnswer containing a text field. All four blocks ship in script tags in the HTML head or at the end of the body; splitting them does not add value.

Validation is non-optional. Schema.org's own validator at validator.schema.org, Google's Rich Results Test at search.google.com/test/rich-results, and the Schema Markup Validator from Yandex each catch a different class of issue; running the HTML through all three before shipping is cheap. The specific gotchas: datePublished and dateModified in ISO 8601, image as an absolute URL and not a relative path, author.url resolving to a reachable page, and publisher.logo as an image with explicit width and height. A schema block that parses but fails Google's richer validation is unlikely to feed Google AI Overviews.

Schema types to skip without a specific reason to add them: standalone HowTo for non-recipe content (Google removed its rich result in 2023), Review or Rating markup without a real reviewed product (fabricated review schema is a documented factor in manual action penalties), QAPage (designed for community-generated question-and-answer sites, not editorial FAQ sections), SpeakableSpecification outside of true voice-answer content, and the long tail of event, course, and product sub-types that do not describe your content accurately. The incremental complexity of adding them generally outweighs any upside, and in some cases triggers structured-data warnings in Google Search Console that distract from the real work.

The measurement gap is specific. Semrush's AI Visibility Index 2026 surveyed 481 marketers and found that 45 percent said they cannot properly measure their brand's visibility in AI-generated answers, and only 9 percent said they can measure all the relevant metrics. The gap is not awareness; it is that classic search measurement (ranking on specific keywords, click-through rate per position) does not transfer to a system where the LLM composes a different answer every time.

Four signals carry the AEO measurement job. First, AI referral traffic by platform in GA4. Set up custom channel groupings that capture referrals from chat.openai.com, chatgpt.com, perplexity.ai, claude.ai, copilot.microsoft.com, and gemini.google.com, with the share subdomains for ChatGPT and Perplexity. Google AI Mode and AI Overviews are not separately attributed in GA4 as of 2026, which is a known gap; traffic from them appears as Google organic unless the user clicked a link in the AI panel specifically, in which case it carries the AI product's referrer.

Second, branded citation share. Query a fixed sample of your 20 to 50 highest-intent target queries on each platform monthly, and record how often your brand or your pages are cited in the answer. The sample is fixed, so results are comparable month over month. Tools that automate this include Profound, Peec AI, SE Ranking's AI Overviews tracker, and Semrush's own AI Visibility feature; the hand-rolled version is a spreadsheet with queries down the rows and platforms across the columns, run on the first business day of each month.

Third, direct-traffic forensics. Direct traffic averaged 31 percent of sessions across the Higoodie panel in 2026, and the report argues that a measurable fraction of it is AI-origin visits with stripped referrers. The practical signal is correlation: when a new AEO content push ships, does direct traffic grow in a time-aligned way with branded citation share? If yes, direct is partly AI, and attributing a conservative share of it as AI-origin produces a more honest total.

Fourth, engagement quality by source. AI-source traffic shows a longer average engagement time (58.5 seconds in the Higoodie panel) than Google organic (44.2 seconds), and a higher engagement rate (67.8 percent versus 61.8 percent). This matters because AEO traffic is not a one-to-one substitute for Google traffic in volume; it is a smaller, more qualified stream. A sales team measuring AEO on raw session count will undervalue it; a team measuring on engaged sessions or on pipeline-sourced metrics will see the real signal. The discipline is the same one applied to [evaluating AI systems in production](/blog/how-to-evaluate-ai-systems): pick the metric that corresponds to the business outcome, not the one that is easiest to generate a dashboard for.

The common mistake is treating AEO measurement like Google SEO measurement. AI engines rarely return the same answer twice to the same query, so position tracking on a specific query is noise, not signal. Citation share across a sample is the right unit.

AEO is not a universal defense. It prevents a specific set of problems and leaves others untouched. Naming them separately is how a team decides what to build and what not to.

What it prevents: being invisible when AI engines synthesize answers in your category, being cited less often than peers on your highest-intent queries, losing a diversified referral mix as the AI platform share redistributes, and leaving the entity graph unanchored while competitors cement theirs. What it does not fix: weak content (a schema-perfect page with vague claims does not get cited over a schema-weaker page with sharp claims), insufficient domain authority (a brand-new site still needs crawl history and backlink growth), factual errors (AI engines punish factual inconsistencies through reputation signals), or a thin crawl budget (if Googlebot cannot cover your site, AI Overviews cannot cite it).

Four mistakes recur in the AEO work we review for US B2B teams. First, writing a different version of the page for AI engines than for readers. The AI engine is reading the same HTML. A page that reads as a marketing pitch to a human reads the same to the LLM, which cites more direct sources instead. The move is to write for both, which is possible because the structural moves (direct-answer leads, FAQ blocks, clear hierarchy) also serve readers.

Second, overbuilding llms.txt. Google's AI optimization guide, published May 15, 2026, states that llms.txt is not used by any Google Search feature including AI Overviews or AI Mode, and groups it with content chunking and AI-specific rewriting as tactics that do not help with Google AI visibility. Anthropic, OpenAI, and Perplexity use it in agent-facing contexts (SDKs, agentic protocols, developer documentation sites). The decision rule: ship an llms.txt file if your site is a target for Claude, OpenAI Agents SDK, or Perplexity agent referrals (developer docs, API references, deep technical content), with a curated list of 10 to 50 links. Skip it for marketing sites with no agent-facing audience, where the time spent is better applied to the first three levers above.

Third, fabricated author schema. An AI engine's entity graph is cross-referenced against the public web; a Person schema with credentials that do not corroborate from a LinkedIn, GitHub, or publication record is a signal that pushes the whole page toward untrusted. For brands publishing as the organization rather than individual authors, the fix is to resolve the author field to the Organization directly; the Organization carries the authority, and the entity consistency work does the lift.

Fourth, treating AEO as a separate initiative from SEO. The Semrush 2026 survey of 481 marketers found that 81 percent of teams with integrated SEO and AI search workflows reported more AI-driven traffic or leads, versus 36 percent of teams running them separately. The same page serves both endpoints; the structural moves that make a page citable also make it rankable. A team that spins up a separate AEO workstream usually ends up with duplicate work and conflicting content calendars. The overlap with the discipline [context engineering for production LLM systems](/blog/context-engineering-production-llm-systems) names at the model-call level is specific: control what the retrieval sees, make the right content legible, strip the noise that compounds against you.

The 60-day plan below is calendar-honest. It is not a full AEO program; it is the first measurable arc that moves citation share from zero or near-zero on a baseline sample toward meaningful presence. Teams running it as a side project take longer; teams with a dedicated owner and executive air cover often finish inside the window.

Weeks one and two, audit. Pick the 20 to 50 queries that would send a qualified buyer to your site. Query each on ChatGPT, Perplexity, Google AI Overviews (where visible in your geography), Claude, and Gemini. Record which of your pages (if any) is cited, which competitor is cited, and which third-party source is cited. In parallel, audit entity consistency: Organization schema served on every page with name, url, logo, sameAs; name consistency across LinkedIn, Crunchbase, press mentions, Wikipedia (if applicable); canonical URL enforcement; and author or publisher presence on every content page. The output is a baseline citation-share table and an entity-signal gap list.

Weeks three and four, schema. Deploy Article (or BlogPosting) plus FAQPage plus BreadcrumbList JSON-LD across every evergreen content page. Validate each page through the schema.org validator and Google's Rich Results Test before merging. Centralize the schema generation in a single library so a schema change ships to every page in one commit. Add the Organization block to the site layout. For non-blog content pages (service pages, pricing pages, offer pages), add the appropriate schema type (Service, Offer, Product) only where the content actually matches the schema's semantics.

Weeks five and six, content retrofit. Take the top 10 highest-intent pages identified in the audit and apply four moves to each. Rewrite the first paragraph to lead with a direct, declarative one-sentence answer to the page's target query. Add an FAQ block with the three to eight questions buyers actually ask, sourced from the audit, from support tickets, and from your sales team's objection list. Confirm the H1 and the first H2 contain the target query's noun phrase. Verify server-side rendering by viewing the raw HTML and checking the article body is present without JavaScript. The retrofit is per-page work; it is not a bulk operation.

Weeks seven and eight, measurement. Deploy GA4 custom channel groupings for AI referral platforms. Set up a monthly branded citation share tracker (hand-rolled spreadsheet or a tool) running the baseline sample against each platform. Set up a direct-traffic correlation watch for AEO-correlated growth. The first comparable month of data arrives at day 30 of measurement running; the second confirms the pattern. Before that, movement in individual queries is noise, not signal.

Where this fits with the rest of your growth stack. AEO is the second endpoint for every page you were already going to publish for SEO. The structural moves (direct-answer leads, FAQ blocks, schema, entity consistency) also lift Google rankings, because Google's own retrieval has moved in the same direction. Stallwart runs AI SEO and AEO engagements for US and UK B2B companies whose Google traffic is holding steady while their referral mix diversifies into AI. The deliverable is the schema library, the retrofit of the top ten pages, the measurement harness that proves the lift, and the ownership path so the team can run the program after the engagement ends. If a team is wondering whether AEO is worth the quarter, the answer is specific to its mix of query intent, its existing content inventory, and its current citation baseline; the audit above is where that answer comes from.