You asked Claude to name the best tools in your category. It named four companies, confidently, and none of them were you.
That stings, and it's fixable. What makes it fixable is understanding something most AEO advice skips: to get cited in Claude, you have to know which of three places the answer came from, because only one of them involves live pages on the web.
This guide walks you through mapping the prompts your buyers ask, auditing what Claude cites instead of you, and building the pages that can win those slots. By the end you'll have a shortlist of specific gaps and a page plan for each one.
What you need: a Claude account, a spreadsheet, and access to your CMS and your robots.txt.
First, know where Claude's answers actually come from
Before you change a single page, spend ten minutes on this. It reorders everything you do next.
Claude composes answers from three pools.
Pool one: what the model already knows. Facts baked into the model during training. As of mid-2026, Claude Sonnet 4.6 trained on data through August 2025 and Claude Haiku 4.5 through July 2025. Anything published after a model's cutoff simply isn't in there. This pool produces answers with no links at all.
Pool two: web search. Claude can decide to run a live search, using Brave Search as the backend. Each search returns a small slate of results, up to ten. When Claude uses this pool, citations show up inline, as links attached to the sentence they support. There's no tidy references list at the bottom.
Pool three: page fetch. A separate tool that reads one specific URL. This is what powers "summarize this article for me." Search finds candidate pages, fetch reads them in full.
Worth saying plainly: Anthropic has not published a ranking methodology for how Claude picks web sources. What follows about Anthropic Claude citations is reconstructed from Anthropic's own documentation on how the tools work, plus practitioner observation of what actually gets cited. Treat the mechanics as solid and the selection signals as well-evidenced patterns rather than published rules.
Here's the part that changes your strategy. Claude decides whether to search, and it usually doesn't bother for evergreen definitional questions. Ask "what is marketing attribution" and you'll likely get a clean, unsourced answer straight from training. Ask "best attribution tools in 2026" and search fires, because the year signals freshness.
So brand visibility in Claude splits in two. Comparison and recency prompts are winnable with content you publish this quarter. Definitional prompts are won slowly, by becoming a name the model absorbed long before anyone asked.
You can influence both. They just move on different clocks.
Step 1: Map the prompts your buyers ask Claude
Start with questions, not pages. Everything downstream gets measured against this list, so it's worth an afternoon.
Build a working set of 25 to 50 prompts a real buyer in your category would type. Spread them across five intents: definitional ("what is X"), comparison ("X vs Y"), recommendation ("best X for Y"), how-to ("how do I X"), and recency ("latest X in 2026"). Then run each one in Claude and record what comes back: a summary of the answer, every URL cited, whether your brand appeared, and which competitors did.
Pro tip: pull the wording from places your buyers already talk. Sales call transcripts, support tickets, Reddit threads, G2 reviews, Search Console queries. If customers say "cheapest tool for X," that's your prompt. "Leading X platform" is your marketing team's prompt, and nobody types it.
You're done when every prompt has an intent label, an answer summary, a cited-URL list, and a yes or no on your brand.
Where people go wrong: optimizing for prompts the team invented in a meeting. A prompt set built from imagination measures imagination.
If staring at a blank spreadsheet is the thing that kills this step, that's the part software can hand you. DeepSmith's AI Visibility module includes Discover Prompts, which generates a starter prompt set from your product, persona, and buyer-stage context. You edit a populated list instead of building one from nothing. Worth knowing up front: DeepSmith tracks ChatGPT, Gemini, Perplexity, Claude, and Google AI Mode, and Claude coverage sits in the Enterprise tier, so check what your plan covers before you count on automated Claude runs.
Step 2: Audit the URLs Claude cites today
Now look at who's winning, not just that you're losing.
Go back through your prompt set and pull out every cited URL. Classify each one twice. First by content type: data study, comparison page, how-to guide, FAQ, listicle, documentation, product page. Second by publisher: your domain, a competitor's, or a third party.
Two patterns tend to jump out fast.
Claude leans hard toward article-shaped URLs. Practitioner analysis of Claude citations found roughly 56 percent of cited URLs sit under a /blog/ path, while homepages and product pages account for around 7 percent. Your beautifully written product page is not the asset competing here. Your blog is.
Claude also cites few sources per answer. Typical responses carry somewhere between one and five. That's a small slate, and established outlets, reference sites, trade publications, and review aggregators take a disproportionate share of it.
You're done when you can name the top five publisher archetypes Claude leans on in your category and roughly what share of citations each one takes.
Where people go wrong: counting only their own citations. The third-party pages Claude cites instead of you are the most useful column in the whole sheet. Those are pages you can get listed on, contribute to, or out-publish.
Step 3: Pick the citation gaps worth closing
You can't fix 50 prompts. You don't need to.
Go prompt by prompt and flag three situations. One, Claude answers with no source at all, which usually means the question resolved from training memory. Two, Claude cites a competitor and never mentions you. Three, Claude cites a third-party page that's genuinely weaker than something you could publish next month.
That third category is the gift. A thin roundup from a site nobody's heard of is a slot you can take.
Narrow to 5 to 15 prompts where a specific new or improved page could plausibly earn the citation. Write the target page next to each one.
You're done when every gap on your shortlist names a prompt, a reason you're missing, and one page that would fix it.
Where people go wrong: picking gaps that are too wide. "Be cited for CRM" isn't a gap, it's a wish. "Be cited for CRM versus spreadsheets for teams under ten people" is a gap, and you can write that page on Thursday.
Step 4: Build pages Claude can extract
This is the step where the work actually happens, and the whole thing turns on one idea: Claude cites passages, not pages.
For each gap, publish in the archetype that's already winning in your category. Usually that means one of these:
- Original research and data studies. Your own survey, benchmark, or internal metrics, published with methodology. Highest authority multiplier of anything on this list.
- Comparison pages. "X vs Y," "alternatives to X," "best X for a specific use case." Side-by-side tables, named criteria, consistent columns.
- How-to guides. Numbered steps, definitions early, named tools throughout.
- FAQ pages. Question as the heading, answer in the first paragraph, no runway.
- Definition and explainer pages. A direct 50 to 80 word answer first, elaboration after. These are your play for the definitional prompts that never trigger a search.
Then write for extraction. This is the concrete meaning of "optimize for Claude answers," and it's mostly formatting discipline. Open every section with a direct answer of roughly 40 to 80 words. Use named entities, named sources, named numbers, and named dates. "47 percent of B2B marketers" is quotable in a way "nearly half of marketers" never will be. Swap pronouns for the actual name in every key claim, so a chunk pulled out of context still makes sense.
Ship it as clean, semantic HTML. Real h1, h2, h3, real lists, real tables, rendered on the server. Claude's training pipeline learned a preference for content it can parse into clean facts, and that preference carries over into how it ranks candidates during live search.
Add schema: Article, FAQ, HowTo, Product, Organization, BreadcrumbList. Here's the honest version of what schema does, though. Ahrefs tracked 1,885 pages adding structured data and found cited pages were roughly three times more likely to carry JSON-LD, while adding schema alone produced no measurable citation lift. Read that as schema helping machines parse you correctly, not as a ranking lever. Add it. Don't expect it to carry the page.
You're done when each priority page answers the question inside the first 80 words, validates clean on structured data, and renders fully without JavaScript.
Common mistake: splitting the answer across tabs and accordions, or burying it under three paragraphs of positioning. If a reader has to click to reveal it, assume the retriever never saw it.
Building pages at this standard repeatedly is where most teams stall, because answer-first structure, schema, headings, and internal links are five separate manual passes per article. DeepSmith's writing pipeline handles those during creation rather than after, using stored brand context so the output is on-voice and product-accurate. It's a production engine, not a first draft you rescue.
Step 5: Layer in the authority signals
Good structure gets you parsed. Authority gets you chosen.
Six signals do most of the work here:
- A named author with credentials, linked to a real bio page. Anonymous posts lose to bylined ones.
- Primary sources cited inline. Link the actual study, dataset, or government page you're drawing from. Citing primaries is a multiplier on your own credibility.
- Visible dates. Published and last updated, both.
- An editorial standards page naming who reviews content and how. Small signal, real one.
- Proof of first-hand experience. "We tested," "we surveyed," "our data." Claude's selection behavior favors pages that show the work.
- A footprint beyond your own domain. Industry publications, review aggregators, awards, press features listed on your About page.
That last one is the one teams skip, and it's the one that hurts most. You will not get cited in Claude on the strength of your own domain alone. Claude rarely cites a brand with no presence outside its own website, no matter how good the on-page content is. Domains carrying institutional weight, including .edu and .gov, hold a measurable premium in most categories, and established publishers take a large share of citations while newer domains accumulate trust.
You're done when every priority page carries an author, a date, inline primary citations, and a path to at least one outlet Claude already cites in your category.
Where people go wrong: treating the About page as decoration. For Anthropic Claude citations, it's an entity record. Make it verifiable.
Step 6: Harden the technical surface
Boring, cheap, and blocking. Do it before you write anything else.
Check four things:
- Server rendering. Priority content has to exist in the raw HTML. A client-side framework with no server rendering can hide your page from the retriever entirely.
- Robots access. Confirm your robots.txt doesn't block Anthropic's crawler and fetch user agents. Plenty of teams added a blanket block in 2023 and never revisited it.
- Sitemap coverage. Every priority URL in it, consistent URL patterns, BreadcrumbList schema where it fits.
- Speed. Fast LCP, clean Core Web Vitals, pages loading in around two and a half seconds.
You can add an llms.txt file while you're in there. Be clear-eyed about it: there's no validated evidence it improves citation rates as of mid-2026. It's cheap hygiene, not a lever. Anyone selling it as one is guessing.
You're done when an audit confirms server-rendered content, an open robots policy for Claude, complete sitemap coverage, and healthy load times on your shortlist.
Where people go wrong: assuming an allow rule proves access. Permission isn't the same as arrival.
Step 7: Refresh on a schedule so recency works for you
Freshness is a live ranking input for exactly the prompts most likely to trigger a search.
Put your shortlist on a quarterly refresh cadence, monthly if your category moves fast. On each pass: update the statistics, swap in current examples, re-verify the underlying data, refresh the visible last-updated date, and add a short changelog line saying what changed.
For recency-driven queries, content published or refreshed in the last 30 to 90 days carries a structural advantage. That's not a trick. It's just that a page dated last week reads as more likely to be right than one dated two years ago.
You're done when every priority page shows a recent last-updated date and its data has been re-checked within 90 days.
Where people go wrong: two versions of the same mistake. Publishing once and walking away, and updating the copy without touching the timestamp. If you fixed it, say when.
Step 8: Measure monthly, then repeat what worked
When you optimize for Claude answers, you're running a loop, not a launch.
Re-run your full prompt set monthly and track four things: mention rate, citation rate, share of voice against your competitor set, and which specific URLs are earning the citations. Those four numbers are what brand visibility in Claude looks like when you stop guessing at it. Then act on what you see. When a page starts getting cited, take it apart and copy its structure into the next gap. When a page drops out, diagnose it: stale data, a competitor who published something better, or a technical regression.
You're done when the prompt set, the scoring, and the refresh calendar live in one place that runs on a rhythm instead of in your head.
Where people go wrong: treating this as a project with an end date. Your category shifts, competitors ship new comparison pages, and Anthropic ships new models with new training cutoffs. Claude assistant AEO is a maintained system, not a launch.
This is also where doing it by hand gets expensive. DeepSmith's AI Visibility Pages view shows which of your pages AI actually cites and what share of your total citations each one carries, which is exactly the input this step needs. It reports what the engines did. It doesn't control what they decide to do next, and nobody's tool does.
What to do next
Pick three gaps from your shortlist. Just three. Publish or rebuild those pages this month.
Then re-run those three prompts in four weeks and see what moved. Claude assistant AEO rewards patience more than volume, and a small set of genuinely extractable, genuinely authoritative pages will outperform twenty thin ones every time.
If your first audit turned up more problems than you expected, that's normal. Almost everyone's does. Most of what you found is unglamorous: a missing byline, an answer buried under three paragraphs, a robots rule someone set years ago. Boring gaps close fast.
If you'd rather not run the prompt set and the page audit by hand every month, that's what a platform is for. DeepSmith gives you tracking and production in one place, starting with a workspace already populated with your brand brief, competitors, and a starter prompt set. Start a free trial and look at your real numbers before you commit to anything. It's 7 days, with no long-term contract.
You're closer than that first audit made you feel. One page at a time.
