Pillar Guide · 2026 Edition · ~30 min read

    The Complete Guide to AI PR

    How B2B tech brands earn citations from ChatGPT, Perplexity, Claude, and Google AI Overviews — and why the discipline that used to be called PR now has to work for both humans and machines.

    By Parry Headrick · Founder & CEO, Crackle PR · Updated

    1. What AI PR actually is (and isn't)

    AI PR is the discipline of earning brand citations from LLM-powered answer engines — ChatGPT, Perplexity, Claude, Gemini, Microsoft Copilot, and Google's AI Overviews and AI Mode — while continuing to earn coverage in the human-consumed press. It sits at the intersection of technical SEO, media relations, thought leadership, and structured data engineering. It is not a rebrand of SEO, and it is not a rebrand of PR. It is the operating model that emerges when both disciplines converge on the same underlying question: what does an AI system trust as a source?

    Three years ago that question was theoretical. In 2026, roughly one in five commercial B2B research journeys begins with an LLM answer rather than a search results page. When a CIO asks Perplexity "who are the leading platforms in data observability?", the brands cited in the response inherit an enormous positioning advantage over the ones that are not. The answer is composed in seconds, treated as authoritative, and — in most cases — used as the shortlist. There is no page two.

    AI PR is what a communications team does deliberately to become one of the brands the machine names first. It is a discipline that has to be run on both the technical and editorial floors of a company, and it has to be measured in a way that neither traditional PR nor traditional SEO teams are set up to measure. That is the discipline this guide is about.

    What AI PR is not

    It is not "just" schema. It is not "just" content marketing at higher volume. It is not "just" a new set of media targets. And it is emphatically not solved by publishing a hundred average blog posts a quarter and hoping the crawler picks something up. AI PR is a systematic discipline. When run casually it produces nothing. When run rigorously it produces one of the most defensible moats a B2B tech brand can build in 2026.

    2. Why the discipline changed between 2024 and 2026

    Between the introduction of ChatGPT's browsing mode in late 2023 and the mainstream adoption of AI Overviews across Google's US SERPs in 2025, the buyer journey for B2B tech quietly rewired itself. Three shifts matter most:

    The retrieval surface exploded. A brand mentioned in a Wall Street Journal feature used to reach roughly the reach of that publication. In 2026 it also reaches every LLM answer that retrieves from that URL for the next several years — every ChatGPT session, every Perplexity query, every AI Overview that stitches the citation into an answer to an adjacent question. The half-life of a tier-one placement effectively went from months to years. But only for brands the LLM can resolve as a single, unambiguous entity.

    Structured data went from nice-to-have to load-bearing. The signals that used to marginally help Google rankings — Organization schema, FAQPage, HowTo, DefinedTerm, canonical URLs — became the primary language LLMs use to understand what a page contains and who published it. A page without schema is not invisible, but it is significantly harder to extract cleanly. A brand without a consistent Organization + Person graph across its properties is difficult to resolve as a single entity at all.

    Attribution broke. Traditional web analytics can identify a referral from a search engine. They mostly cannot identify a referral from an LLM answer — the user reads the answer, forms an opinion, and eventually navigates directly, via branded search, or via a channel that carries no source attribution. This means the entire measurement toolkit that PR and SEO teams have relied on for a decade produces a systematic under-count of AI-driven demand. The dark-funnel problem, which used to be a rounding error, is now the dominant measurement challenge.

    Individually, each of these shifts would have been survivable. Together, they force a new operating model. That model is what the rest of this guide describes.

    3. How LLMs actually choose sources

    Every major LLM answer engine now combines pre-trained knowledge with real-time retrieval — a pattern known in the literature as retrieval-augmented generation, or RAG. The exact mix varies by model and by query type, but the mechanics converge:

    When a user asks an LLM a question with commercial or factual weight ("what are the best cybersecurity vendors for financial services?"), the system decomposes the question, issues retrieval queries against a curated index of high-authority sources, ranks the retrieved passages by relevance and source authority, and composes an answer that stitches together the top-ranked extractions. The citations shown to the user are pulled from the passages the model actually used — not from a separate ranking layer.

    Three things follow from this that most PR teams underestimate.

    First, source authority is not the same as Google's authority score. LLM retrieval indices weight a smaller, curated corpus more heavily — tier-one national press, major trade publications, established analyst firms, .edu and .gov domains, Wikipedia, and a long tail of category-specific outlets. A domain with excellent Google authority but no presence in that curated corpus is comparatively invisible to the LLM. A domain that is comparatively small on Google but well-cited by the curated corpus punches well above its search-visibility weight.

    Second, passages beat pages. LLMs retrieve extractable passages, not URLs. A three-thousand-word article without a clean summary, without structured lists, without labelled quotations, and without schema markup is harder for the model to extract cleanly than an eight-hundred-word article with a labelled TL;DR, a five-bullet summary, and FAQPage schema. This is why the same story, told twice with different structure, produces very different AI citation outcomes.

    Third, entity resolution is the gate. When the model composes an answer, it needs to attribute the extraction to a specific entity — a company, a product, a person. If your brand is referred to inconsistently across the sources the model retrieves (different capitalisation, different framing, no shared entity graph), the model will either drop the attribution or attach it to a competitor whose entity is clearer. Entity clarity is not a nice-to-have; it is the mechanism by which citations attach to your brand rather than someone else's.

    “LLMs don't cite pages. They cite entities. If your brand isn't resolvable to a single entity across the sources the model retrieves, you can be quoted extensively and still not appear in a single answer.”

    Parry Headrick, Founder, Crackle PR

    4. The AI Citation Pyramid: the operating model

    Everything in this guide operationalises through a single framework we developed at Crackle PR: the AI Citation Pyramid. Five stacked layers, each dependent on the ones beneath it, each producing a different kind of signal. The full explainer lives on the framework page; the shorthand for this guide is:

    1. 5

      Tier-one validation

      Earned coverage in outlets LLMs weight most heavily. This is the apex — the layer that compounds hardest.

    2. 4

      Distributed proof

      Referring-domain diversity from credible sources. The leading indicator of everything above it.

    3. 3

      Owned depth

      Pillar pages, named frameworks, original data. This is where you become extractable.

    4. 2

      Structured facts

      Glossary, FAQ, HowTo, DefinedTerm schema, machine-readable indices. This is the language the model reads.

    5. 1

      Entity clarity

      Organization + Person schema, canonical URLs, consistent sameAs cluster. The foundation.

    The pyramid is prescriptive, not descriptive. You cannot skip layers. A brand at layer 5 with a broken layer 1 will consistently under-perform a brand at layer 3 with a rock-solid layer 1, because the LLM cannot attach the tier-one citations to a resolved entity. This is the single most consequential idea in this guide.

    The pyramid also tells you where to spend. If your AI Visibility Score shows red at layers 1–2, spend nothing on layer 5 until those are fixed. If layers 1–3 are green and layers 4–5 are red, you have infrastructure without amplification — the fix is earned media, not more schema. Score honestly, then sequence accordingly.

    How the other Crackle frameworks slot in

    The pyramid describes the visibility stack. Three sibling frameworks describe the disciplines that populate it:

    • The Earned Media Flywheel is the operating model for layers 4–5 — how a single tier-one hit gets amplified, captured, attributed, and re-invested into the next cycle.
    • The Reputation Capital Model is the balance-sheet view of what you're accumulating. AI citations are a compounding reputation asset — arguably the highest-quality kind — but they compound only if you're building reserves during calm periods, not just when you have a launch to promote.
    • The Category Velocity Engine is the play for brands whose product doesn't fit an existing category. Owning a category name in LLM answers is the strongest possible AI moat, because the category becomes an entity that gets cited alongside your brand every time it's queried.

    You do not need all four frameworks operating at once. You do need to know which one is doing the most work at each stage of your program.

    5. Execution: the twelve plays that produce citations

    The pyramid is the model. The plays are what actually gets done. In our engagements, twelve plays account for roughly 80% of measurable AI citation lift. They are ordered by layer, not by priority — start from the bottom.

    Layer 1 — Entity clarity

    Play 1: Publish a canonical Organization + Person JSON-LD graph on the homepage and key money pages, with sameAs links to every property that carries the brand — LinkedIn, Crunchbase, Wikipedia (if listed), press mentions, GitHub, industry associations. Validate every page through Google's Rich Results Test. This is the anchor the entire pyramid is measured against.

    Play 2: Canonicalise ruthlessly. One brand name spelling, one URL per canonical page, one preferred domain (www or apex, not both). Redirect all variants. LLMs treat a brand with ambiguous canonicalisation as multiple weak entities rather than one strong one — a common cause of underperformance that is invisible to normal SEO reporting.

    Layer 2 — Structured facts

    Play 3: Build a glossary of every proprietary and category term you use, each on its own URL with DefinedTerm schema. Aim for 100–300 entries over 12–18 months. Glossary pages are among the highest-density citation targets because they answer the exact question the LLM is trying to satisfy.

    Play 4: Add FAQPage and HowTo schema to every substantive page, with speakable markup on the answer bodies. Not because voice matters (it mostly doesn't yet), but because speakable is a signal to the model that this passage is designed to be extracted cleanly.

    Play 5: Publish a machine-readable index. An llms.txt or /.well-known/ endpoint that describes your site to AI agents; an MCP server if you have the engineering appetite. This doesn't move the needle on retrieval today, but it puts you ahead of the crawler-policy shifts that are already visible on the horizon.

    Layer 3 — Owned depth

    Play 6: Publish 3–5 pillar pages of 3,000+ words each, answering the full category question end-to-end. This guide is a pillar page. Pillar pages consolidate topical authority in a way that a hundred short posts cannot, and they are the primary asset LLMs cite when the query is broad.

    Play 7: Publish named, proprietary frameworks on dedicated URLs. Named concepts are cited disproportionately because they resolve ambiguity for the model. Every framework you name and defend is a stable entity the LLM can attach to your brand. Trademark them if the play is important.

    Play 8: Publish original data on a recurring cadence. Quarterly indices, benchmark reports, surveys. Original data is the single highest-quality citation asset because it cannot be paraphrased away — the LLM has to attribute the number. Our AI Visibility Index is the canonical example.

    Layer 4 — Distributed proof

    Play 9: Get founder-led thought leadership into circulation monthly, through bylines, LinkedIn essays, podcast appearances, and quoted commentary. Referring-domain diversity is the leading indicator for everything at layer 5; the founder is usually the most productive mechanism for producing it.

    Play 10: Syndicate the data. A quarterly benchmark that lives only on your owned domain reaches only your existing audience. A quarterly benchmark that is picked up by ten trade publications and quoted in one national outlet reaches the retrieval corpus itself.

    Layer 5 — Tier-one validation

    Play 11: Earn 3–6 tier-one placements a year, each engineered for AI extractability. This means quoted attribution to a named spokesperson, structured claims a journalist can lift into a summary, and a linked reference to a canonical URL on your domain. Traditional press coverage without those attributes still helps humans; AI PR requires the added structure.

    Play 12: Systematically pursue analyst inclusion. Gartner, Forrester, IDC, and category-specific analyst firms. Analyst research is disproportionately cited by LLMs because it is treated as authoritative by the retrieval corpus. Analyst relations is the highest-effort play in the twelve, and it is worth every hour.

    6. Measurement: KPIs, proxies, and the dark-funnel problem

    Measurement is where most AI PR programs fail politically, if not operationally. The instinct is to report the same metrics traditional PR reports — placements, impressions, share of voice — and add a "citations" column. This produces reports that look reasonable, mislead the board, and collapse under scrutiny within two quarters. The honest measurement stack looks different.

    Layer 1 metrics: entity coverage. Number of canonical pages carrying valid Organization + Person schema. Number of properties in the sameAs cluster. Number of validation warnings from Google's Rich Results Test. These are hygiene metrics; they should trend to zero errors and stay there.

    Layer 2 metrics: extraction surface. Number of glossary entries live. Number of pages with valid FAQPage / HowTo / DefinedTerm schema. Number of terms defined on canonical URLs (vs paraphrased across many pages). These grow monotonically and act as the base multiplier for every downstream tactic.

    Layer 3 metrics: owned depth. Word count of published pillar content. Number of named frameworks live. Recency of original data. This is where the composition of your top 25 pages matters more than the count of your top 250 pages — deep, cite-worthy resources compound; thin ones do not.

    Layer 4 metrics: referring-domain diversity. Monthly delta in referring domains from credible sources. Share of referring domains that are in the LLM retrieval corpus (as best you can approximate it via tier-two-and-above trade press, analyst firms, and industry associations). This is the leading indicator most PR teams underuse.

    Layer 5 metrics: citation share of voice. Number of tracked AI queries where your brand appears as a cited source, split by engine (ChatGPT, Perplexity, Claude, Google AI Overviews). Share of voice against category competitors on the same queries. This is what Semrush's AI Search reporting, Otterly, Profound, and similar tools attempt to measure. Treat the absolute scores as directional; the composite trend and the competitive comparison are the honest signals.

    The dark-funnel problem, and how to work around it

    Even with all of the above in place, most of the demand AI PR generates is not directly attributable. A CFO reads an answer that cites your brand, forms an opinion, and eventually arrives via direct traffic, branded search, or a channel that carries no source attribution. The lift is real; the attribution is broken.

    The workable proxies are:

    • Direct-traffic and branded-search delta, quarter over quarter, correlated against citation-count growth. When citation growth precedes direct-traffic growth by 30–60 days, the causal story is defensible even without click-level attribution.
    • AI-referral session counts where they exist. Some engines pass a referrer header; most do not. A lightweight client-side detector plus consistent UTM tagging on any URL published inside AI-answer-friendly assets catches the sliver of clickthrough that is attributable. It won't be most of the funnel, but the ratio of tagged sessions to total citations over time is a useful compression signal.
    • Lead-source self-report. Add "How did you first hear about us?" as a free-text field on every high-intent form. Manually tag the responses monthly. This produces a slow but qualitatively honest read on what the analytics cannot see.

    7. The seven mistakes that cost brands AI share of voice

    Common, expensive, and — mostly — fixable:

    1. Chasing tier-one placements on a broken entity graph. A great WSJ hit on a brand the LLM can't resolve produces no citation lift. Fix layer 1 first.
    2. Publishing volume over depth. A hundred average blog posts underperforms fifteen deep pillar pages every time. The pillar pages become the entities the LLM cites; the thin posts become noise.
    3. Not naming your frameworks. Unnamed concepts get paraphrased; named concepts get cited. The naming decision is a marketing decision, not a copywriting decision.
    4. Treating AI PR as a channel instead of a program. A campaign runs for six weeks and stops. A program runs for six years and compounds. AI citations reward programs.
    5. Measuring citations without measuring referring domains. Citations are the outcome; referring domains are the input. If you're not tracking the input, you cannot diagnose the outcome.
    6. Under-investing in the founder. The founder is usually the most productive spokesperson, the highest-conversion LinkedIn account, and the fastest path to referring-domain growth. Treating the founder as a nice-to-have is a strategic error.
    7. Reporting to the board on placements. Boards ignore placement lists. They respond to citation share of voice, direct-traffic delta, and pipeline-attributable lift. The reporting format determines the seriousness of the discipline.

    8. Organising the team: who owns AI PR internally

    AI PR sits between three functions — comms, marketing, and the technical / SEO team — none of which is naturally set up to own it end-to-end. The workable configurations we see:

    Small teams (≤5 marketers): comms owns the program, contracts the technical base. Head of comms (or the CEO in a founder-led company) owns editorial calendar, spokespeople, narrative, and earned media. Schema, glossary infrastructure, and pillar-page engineering are contracted to an agency or freelance specialist. This is the fastest path to a running program.

    Mid-sized teams (5–25 marketers): dedicated AI visibility lead, embedded in comms with a dotted line to SEO. Owns the composite metrics, owns the pyramid, briefs both editorial and technical work. This is the configuration that produces the best results in our client base.

    Large teams (25+): a small AI PR pod inside comms. Two to four people: an editorial lead, a data lead (for the recurring benchmark work), a technical lead, and a measurement analyst. The pod runs the pyramid end-to-end and consults into both the traditional PR and SEO teams. Rare at this scale in 2026; more common by 2028.

    What does not work: assigning AI PR to the SEO team and hoping they'll figure out earned media, or assigning it to the PR team and hoping they'll figure out schema. Both fail predictably and politically. The role has to sit with whoever can own both the editorial and technical threads, and that person almost always has a communications background rather than a search background.

    9. Budgeting: what AI PR should cost in 2026

    Ballpark, for a series B–C B2B tech company aiming to reach the "Established" tier on the pyramid within twelve months:

    • Foundational build (one-time or first quarter): $30–80K for schema implementation, glossary infrastructure, canonical clean-up, initial pillar pages, and named framework build. Higher end includes original data collection and a first benchmark.
    • Ongoing program (monthly retainer): $12–45K depending on volume. At the lower end you get the editorial cadence and measurement layer, with in-house engineering handling technical execution. At the upper end you get end-to-end program ownership including tier-one media relations, analyst relations, and quarterly data refreshes.
    • Tier-one earned media (variable): Effectively baked into the retainer for senior-led agencies. If separated as a project fee, expect $15–40K per tier-one exclusive depending on complexity and the required embargo choreography.
    • Analyst relations (annual): $50–150K for a serious program, including analyst briefings, subscription fees, and the internal time required to respond to inquiries and RFIs.

    The pattern we see repeatedly: companies that under-invest in the foundational build spend the same total amount over eighteen months and produce a fraction of the citation growth, because every month of earned-media spend at the top of the pyramid is dampened by a leaking base. The sequencing is what determines the ROI, not the aggregate budget.

    10. Frequently asked questions


    Where to go from here

    If this guide is the map, the frameworks are the terrain and the tool is the compass. In order:

    Talk to Crackle PR

    We run the pyramid end-to-end for a small number of senior-led engagements each quarter.

    Scope a program
    Published by Parry Headrick · Founder & CEO, Crackle PR · Updated
    12
    The Inquiry

    Start with the best tech PR agency.

    Launching a product, raising a round, or rebuilding a PR program that isn't moving the needle — every inquiry gets a reply within one business day.

    (415) 246-8486[email protected]
    Boston · Portsmouth, NH · San Francisco
    Monday–Friday · 9:00–18:00 ET

    A note from the founder

    "We answer every inquiry personally. No bots, no junior staff, no triage queue. If it's a fit, we'll tell you. If it isn't, we'll tell you that, too."

    Parry Headrick · Founder

    By submitting, you agree to be contacted by Crackle PR. We never share inquiries.