AI search doesn't rank you. It quotes you.

Chapter 01

How an AI answer actually gets built

Traditional SEO optimizes for ten blue links. AI search has no list. A model takes a question, silently rewrites it into a handful of sub-questions, retrieves passages from a small pool of pages, and synthesizes one answer with a short source list attached.

That changes what you optimize for. You're no longer competing for position — you're competing to be retrieved, quoted, and attributed. A page can rank #4 in Google and never be cited, and a page that ranks nowhere can be cited constantly because it answers a sub-question cleanly in one extractable paragraph.

The fan-out is the real search

When someone asks a buying question, the engine doesn't search that string. It fans it out. Here's what a single vendor-selection prompt typically becomes behind the scenes:

Three rules that follow from this

1. Retrievability beats rankings. Your content has to be parseable into self-contained passages. A model quotes paragraphs, not pages.
2. Coverage beats depth. Every sub-query without a matching page is a gap a competitor fills. Ten precise pages outperform one 5,000-word guide.
3. Corroboration beats claims. Models weight what multiple independent sources say about you. Your own site alone rarely carries a recommendation.

Chapter 02

Baseline your visibility before you change anything

You cannot improve a number you have never measured. Before any technical work, build a prompt set and score yourself. This takes an afternoon and it will reframe your entire content plan.

Step 1 — Build a 40-prompt buyer set

Write the questions your buyers type into ChatGPT in the two weeks before they shortlist vendors. Five categories, roughly eight prompts each:

• Category: "best [category] software for [segment]"
• Comparison: "[you] vs [competitor]", "alternatives to [market leader]"
• Problem-first: "how do we handle [painful workflow] without [manual process]"
• Commercial: "how much does [category] cost", "is [competitor] worth it"
• Trust: "is [you] legit", "[category] compliance requirements"

Step 2 — Run each prompt across four engines

ChatGPT, Perplexity, Claude, Gemini and AI Overviews. Log three things per run: were you mentioned, were you cited with a link, and which domains appeared instead. Mentions without links still shape the recommendation, but citations are what drive traffic.

Step 3 — Score citation share

Citation share is your citations divided by all citations across the set. It's the one number worth reporting to a board, because it's directly competitive:

Chapter 03

The technical foundation

This is step one of my process for every client, and it is where most companies quietly lose before they start. AI crawlers are less forgiving than Googlebot. If your content isn't in the raw HTML and your bots aren't allowed in, nothing else in this guide matters.

Let the right bots in

Many sites block AI crawlers by accident — through a CDN bot-fighting rule, a security plugin default, or a robots.txt copied from a blog post about "protecting your content from AI." Check your logs for these user agents and confirm they're getting 200s:

# robots.txt — allow the crawlers that feed answer engines
User-agent: GPTBot          # ChatGPT training + browsing
User-agent: OAI-SearchBot   # ChatGPT search index
User-agent: ChatGPT-User    # live user-triggered fetches
User-agent: PerplexityBot
User-agent: ClaudeBot
User-agent: Google-Extended # Gemini / AI Overviews grounding
User-agent: Bingbot         # still powers several engines
Allow: /

Sitemap: https://example.com/sitemap.xml

Server-render everything that matters

Most AI crawlers don't execute JavaScript the way Googlebot does. If your pricing table, feature list or FAQ is injected client-side, it's invisible to retrieval. Run curl -A "GPTBot" https://yoursite.com/pricing and read what comes back. If your key facts aren't in that output, fix rendering before writing another word of content.

Structure the page so passages can be lifted

Schema that engines actually use

Schema won't manufacture authority, but it removes ambiguity about what your entity is. Prioritize four: Organization (with sameAs pointing to every profile you control), Product/Service, FAQPage, and Article with a real author that has its own page.

// Organization — the entity anchor
{
  "@context":"https://schema.org",
  "@type":"Organization",
  "name":"Example Inc",
  "description":"Contractor payroll for distributed teams in 190 countries.",
  "foundingDate":"2019",
  "sameAs":[
    "https://www.linkedin.com/company/example",
    "https://www.g2.com/products/example",
    "https://www.crunchbase.com/organization/example"
  ],
  "knowsAbout":["contractor payroll","global compliance","multi-currency payouts"]
}

Publish an llms.txt

A plain-text map at /llms.txt naming what you do and which page is canonical for which topic. Adoption is uneven, but it costs an hour and forces you to articulate your entity clearly — worth it regardless.

# Example Inc
> Contractor payroll and compliance for distributed teams in 190 countries.

## Core pages
- [Pricing](https://example.com/pricing): published per-contractor pricing, no quote required
- [vs Deel](https://example.com/vs-deel): feature and pricing comparison, updated monthly

Chapter 04

Content that gets quoted, not just read

Step two: map the pages your buyers search at the moment of decision, and build them to be extracted. Traffic isn’t the goal. Being the sentence the model repeats is the goal.

The five page types that earn the most citations

Write the answer first

Structure every section so the first two sentences under a heading are a complete, self-contained answer. Everything after is supporting detail for humans. This single habit moves extraction rates more than any other writing change.

Make every passage stand alone

Retrieval chops your page into chunks. A chunk that says "as mentioned above, it also supports this" is unusable. Repeat the subject. Repeat the qualifier. Slight redundancy reads fine to humans and is what makes a passage safe for a model to quote.

Show your work with a demonstration

Here is the same fact written two ways, and what happens to it:

Chapter 05

Off-site authority: the half nobody does

Step three. Models corroborate. When a buyer asks who's best, engines lean on sources with no stake in the answer — review platforms, publications, roundups, communities. If your name only appears on your own domain, you won't be recommended, however good your pages are.

Where answers about vendors actually come from

• Review aggregators. G2, Capterra, TrustRadius and their niche equivalents are cited constantly in vendor questions. A steady drip of recent, detailed reviews is a citation channel, not a vanity metric.
• "Best of" listicles. Someone else's roundup is often the retrieved source. Pitch to be included; supply the writer with a clean, factual one-paragraph description and current pricing so they quote you accurately.
• Industry publications. Bylined expert features on sites buyers already trust. This is the work that moves both link authority and entity recognition at once.
• Communities. Reddit and specialist forums appear in a striking share of B2B citations. You can’t fake this — but you can participate honestly where your category is discussed.

The entity consistency check

Models build a picture of your company by reconciling everything they've read. If your positioning reads differently on your homepage, LinkedIn, G2 and Crunchbase, that picture is blurry — and blurry entities don't get recommended. Write one 25-word description of what you do and who for, then make it identical everywhere.

// One sentence. Same words. Every surface.
Homepage meta  →  "Contractor payroll and compliance for distributed teams in 190 countries."
LinkedIn about →  "… for distributed teams in 190 countries."
G2 / Crunchbase → "… for distributed teams in 190 countries."
Schema desc    →  "… for distributed teams in 190 countries."

It feels almost too simple. It is also one of the fastest wins available, and it's free.

Chapter 06

The 30-day refresh — why I never publish and forget

Step four, and the least intuitive one. Answer engines strongly favour recent content. Research on citation patterns puts about 76% of AI citations on content updated within the previous 30 days. That single fact should restructure your content calendar.

Most teams spend the whole budget on new posts. The higher-return move is a monthly refresh cycle on pages that already have a chance of being cited.

What a real refresh includes

Changing the date is not a refresh, and engines aren't fooled by it. A refresh that works touches substance:

1. Update every statistic and price, with the source and date named in the text.
2. Add one genuinely new section answering a sub-query the page didn't cover.
3. Rewrite the opening answer so it's tighter and more quotable than it was.
4. Refresh competitor details — their pricing changed, and pages that reflect that get cited over ones that don't.
5. Add an explicit "Last updated: [Month Year]" line in visible body text, not just in schema.
6. Re-submit in Search Console and update dateModified.

Chapter 07

Measurement: proving it in revenue, not rankings

If your reporting can't be connected to pipeline, you're measuring activity. Here are the six views I build for every client, and the tracking underneath them.

Setting up AI referral tracking

Most analytics setups bucket AI traffic into "direct" or "referral" and lose it. Fix that first, in GA4:

• Create a channel group for AI sources matching chatgpt.com, perplexity.ai, claude.ai, gemini.google.com, copilot.microsoft.com.
• Segment AI traffic against organic on what matters: pages per session, demo requests, pipeline created.
• Ask "how did you hear about us?" on your demo form. Self-reported attribution catches what analytics can't, because a lot of AI-influenced buyers arrive by typing your name into a browser.

Chapter 08

The 90-day plan

Everything above, sequenced — roughly how a client engagement runs, and the order to follow if you're doing it yourself.

Chapter 09

Nine mistakes that cost companies citations

1. Blocking AI crawlers by default. Usually a CDN rule nobody remembers enabling.
2. Client-side rendering the facts. If pricing loads in JavaScript, it doesn't exist to retrieval.
3. Hiding pricing. "Contact sales" removes you from every cost-related answer in your category.
4. Writing for word count. A 4,000-word guide with no extractable passage loses to a 600-word page that answers one question.
5. Refusing to name competitors. If you don't write the comparison pages, an affiliate site writes them for you.
6. Publishing once and moving on. Freshness matters in answer engines in a way it never did in Google.
7. Ignoring off-site. On-page work with no corroboration gets you retrieved but not recommended.
8. Chasing volume keywords. The valuable prompts are long, specific and low-volume.
9. Measuring sessions. AI traffic is small and high-converting. Report pipeline or you'll kill the channel before it compounds.

Chapter 10

Proof this works

Numbers from client analytics and citation tracking. The playbook above is what produced them.

If you'd rather not run it yourself

What I do, in five steps

One senior operator, full-stack execution, no outsourcing, no account manager in the middle. The same sequence you just read, run weekly.

Questions

FAQ