For a decade the job was simple: build a fast site, rank it on Google, wait for clicks. Your website was the destination — where a prospect landed, read, and decided. Every dollar of content and SEO spend assumed that finish line.

That assumption broke. A large, fast-growing share of buyer research now happens inside an AI answer — Google's AI Overviews and AI Mode, ChatGPT, Perplexity, Copilot, Claude. The buyer asks, reads a tidy synthesized response, and often acts without visiting anyone's site. By the time they reach you, the shortlist is already set. The category has been explained to them; a few brands have been named as the credible options.

You've seen vendors, Sitecore included, racing to add "AI visibility" tooling and talking about writing "for machines so humans can experience it." That's a symptom, not the cause. The cause is bigger and simpler: the company that owns most of the world's search — Google — rebuilt how search works. Everyone else is adapting to the rules Google just rewrote. So that's where a serious strategy begins.

To fix your visibility you first have to see what the machine actually does when someone searches. It's less mysterious than it sounds — and once it's clear, the whole strategy falls into place.

From ten blue links to one written answer

Old model: Google matched your keyword, ranked the pages, handed over ten links. The user did the hard part — opening tabs, comparing, synthesizing. Ranking near the top roughly guaranteed traffic, so SEO was about climbing that list. New model: the engine does the synthesis for the user. It reads across sources, writes one answer, attaches a few citations. Your goal is no longer to rank a page — it's to be a source the answer is built from.

"Query fan-out": one question becomes a dozen

Here's the mechanic that broke keyword-era marketing. When someone asks a full question, the engine doesn't hunt for that phrase. It behaves like a diligent research analyst: it brainstorms the smaller questions it would need to answer to satisfy the person, then answers them all at once. Google calls this query fan-out — one question split into roughly 8–12 sub-queries, searched in parallel, woven into a single answer.

Three flavors of the same shift

AI Overviews are the summaries that appear automatically atop Google results — in categories like healthcare, education, and B2B tech they now show on the majority of searches. AI Mode is a dedicated conversational tab with deeper reasoning and "deep research" that fires hundreds of sub-queries. And agentic search is where this heads: the engine stops answering and starts acting — comparing, planning, even buying on the user's behalf. In that world, an agent evaluates you before a human ever does.

The numbers that should reset the boardroom

The plain conclusion: AI visibility is a distinct discipline that sits on top of SEO — not a free byproduct of it.

What Google itself says to do (and skip)

Google's 2026 guidance, de-hyped: the fundamentals still rule — useful, people-first content; crawlable, indexable pages; clear structure; real credibility. For its surfaces you don't need special "AI-only" files, and Google warns against thin one-page-per-question sprawl and fake mentions. Build one comprehensive, well-linked resource that walks the buyer's whole path instead. Two caveats: that "you don't need special files" advice is for Google — other engines (ChatGPT via Bing, Perplexity, Claude) have their own logic where structured data still helps — and "it's still SEO" understates how differently citation behaves from ranking, which is exactly why you measure AI surfaces separately.

"Be visible in AI" is a slogan, not something you can manage. To make it real, know exactly what you're judged on: five surfaces and six signals.

Five surfaces, five different rulebooks

There's no single "AI." There are five retrieval systems, each picking sources differently — winning one doesn't win the others. The overlap between what one engine cites and another cites is small, which is why a single blended "AI score" misleads.

Six signals — a ladder you climb in order

Fail a lower rung and the higher ones never matter; the engine never reaches them.

And the floor beneath all of it is still SEO. A slow site, thin content, or weak authority won't be cited no matter how neat your FAQ markup. AI visibility is a new layer on solid foundations — not a replacement.

The work is a simple, repeatable discipline. Identify where you're absent, misrepresented, or out-cited. Bridge the gaps by fixing content, structure, and technical signals. Validate that it worked — per surface — and feed results back in.

The crucial word is loop. AI answers shift constantly — models refresh, competitors publish, facts go stale — so a page cited this month can drop next month. One-and-done is a false finish line. Winners run this as a standing rhythm: priority topics monthly, the full set quarterly. How you run it depends on your platform — which brings us to the two paths.

Where you stand today changes your speed and automation — not your destination. Path A (SitecoreAI / XM Cloud) runs the loop as governed AI agents inside one system. Path B (classic XM/XP) runs the same method as a portable toolkit you assemble. Both reach a brand that shows up, accurately, in AI answers.

Identifying gaps comes down to running a set of prompts and sorting the results. The prompts are the reusable asset. On Path A they become AI agents; on Path B, a prompt library plus "skills." Because Path A's agents just wrap these same prompts, Path B is also your fallback: when an agent isn't set up, run the prompt by hand and lose nothing but speed.

The four kinds of gap

Every answer you test falls into one bucket, and each needs a different fix — which is why you classify before you act:

• Absence — you don't appear. A content or authority gap.
• Misrepresentation — you appear, but it's wrong. An entity-consistency gap.
• Out-cited — a competitor appears instead. An extractability or authority gap.
• Uncited-but-present — mentioned, no link. A structure gap: you're in the answer without the credit.

The prompt library

The reusable prompts. Fill the {{...}} slots once with your brand, competitors, and segments — no coding required. Paste into any AI assistant, or hand them to your team as standard operating scripts.

# Generate the buyer-query probe set
You are a demand researcher. From this positioning:
  brand = {{BRAND}} ; category = {{CATEGORY}}
  icp = {{SEGMENTS}} ; competitors = {{COMPETITORS}}
Produce 40 questions a buyer would ask an AI assistant, covering:
category discovery, "X vs Y", problem-first, capability, pricing/fit,
and trust/proof. Full questions, no keywords. Numbered list.

{{BUYER_QUESTION}}

# capture per surface: full answer; was {{BRAND}} mentioned? cited
# with a link? which competitors appear, in what order? source URLs?

Here is an AI answer to "{{QUESTION}}": """{{ANSWER}}"""
Return JSON:
{ "brand_mentioned": true|false, "brand_cited": true|false,
  "competitors_ranked": [...], "inaccuracies": [...],
  "gap_type": "absence" | "misrepresentation"
             | "out_cited" | "uncited_present" }

For the query "{{QUESTION}}", generate the 10–12 sub-queries
an AI engine would fan out into. For each, tell me which of our
pages ({{URL_LIST}}) answers it — and flag every sub-query
we have zero coverage for.

Given these classified gaps: {{GAPS_JSON}}
Score each 1–10 on (a) buying intent and (b) winnability (how weak
the cited source is). Rank by (a × b). For the top 15, write a brief:
target question + fan-out sub-queries to cover, an answer-first
passage (130–170 words), the table/list/stat to add, schema type,
author signals, and internal links.

You already have the platform this capability is being built into. Put it to work.

Identify

Run the engine above as agents in Agentic Studio: generate 100–300 probes, fire them across all five surfaces, and produce a ranked, typed gap backlog sorted by what's both valuable and winnable. Lean on the fan-out coverage check — the sub-questions you don't answer are where citations quietly leak.

Bridge

This is where XM Cloud earns its keep, because fixes happen where the content lives. With a human approving, use agents to restructure priority pages so each answer is easy to lift: a clear opening definition, a short answer-first summary near the top, a comparison table and a list, specific facts, and headings phrased as the questions buyers ask. Ship structured data — machine-readable labels (JSON-LD) that say "this is a product, this is an FAQ, this is the author" — through your rendering or edge layer. Publish a small llms.txt pointing crawlers to your best pages. Deliberately allow the AI crawlers you want. Reconcile the facts about your brand everywhere they live so the model gets one story. Then wire the loop: gap → brief → drafted fix → human approval → publish → next-cycle check.

Validate

Use AI-search tracking plus the scorecard below. Path A's edge: measurement and action live in one system, so you can attribute a citation lift to the exact change that caused it, and watch your share of the answer grow topic by topic.

No migration required. Attach the same six signals to your current platform and own the connections an all-in-one system would automate. It's more hands-on — that's fine.

Identify

Run the same prompt library as "skills" on a cadence: a documented manual panel (free, closest to ground truth) plus a third-party AI-visibility tracker to scale it. Trust the per-question, per-surface detail over any aggregate score. You control the infrastructure — use it: check server/CDN logs to see whether AI crawlers even reach you, and analytics for visitors arriving from ChatGPT or Perplexity. Zero crawler visits is itself a finding.

Bridge

Start with the step most often broken on legacy sites: get crawlable and indexed. Allow the AI crawlers in robots.txt, and — the highest-return, most-overlooked move — submit your sitemap to Bing Webmaster Tools, because ChatGPT and Copilot lean on Bing and most teams have only ever used Google's. Add machine-readable structured data without a rebuild: emit it from a shared component, generate it from existing content fields via a pipeline step, or inject it at the CDN edge for locked-down sites. Build one "citation-ready" page pattern — definition-first intro, summary block, comparison table, FAQ, key-facts module, named authors, visible dates — and roll it to your highest-value topics first. Publish llms.txt, and reconcile brand facts across Wikipedia, Wikidata, review sites, and your own boilerplate. If migration isn't near-term, a lightweight headless or edge front-end over your existing content — just for pages that matter most — buys modern control over speed and structure without a big-bang replatform.

Validate

Run the same scorecard. Your loop is stitched, not closed, so name an explicit owner for the hand-off from "we measured a gap" to "it's in the backlog." That unowned seam — not any technical limit — is the most common reason Path B programs stall.

Your answer to "are we actually available to AI?" — the same instrument for both paths. Golden rule: score every line per surface and always show the trend, never one blended number. A rise on Perplexity while ChatGPT stays flat isn't noise — it's a diagnosis (probably a Bing-indexing problem) a blended score would bury.

Governance that keeps this safe

A few guardrails regardless of path. Keep a human approving every publish, whether the draft came from an agent or an editor — it protects brand voice and accuracy. Treat allowing AI crawlers as a real decision made with legal and brand, not an accidental default, because it enables both citation and the use of your content. Watch accuracy continuously, not quarterly: a model confidently stating something false about your pricing or product is a reputational risk you want to catch immediately. And keep structured data and brand facts to public, accurate information.