Get a free AI visibility report
.png)
SEO for Claude is a different job from ranking in Google, because Claude answers a disproportionate share of work questions. The shortlist it produces rarely matches Google's results, and this guide covers how to change that.
Most teams discover the problem the same way. A prospect mentions that they asked Claude for options in the category, then lists three competitors and not you. Nobody changed your rankings. Nobody wrote a negative review. You simply were not in the pool of sources Claude reached for. Worse, there is no SERP to check afterward.
That gap is what SEO for Claude is meant to close. The work overlaps with traditional search optimization in places, since crawlability and content quality still matter. Yet the decision logic sits somewhere else entirely. This guide walks through six steps that move a brand from absent to cited. Each step covers the technical checks, content patterns, off-site work, and measurement involved.
Primary lens: model divergence. Every platform assembles answers from its own retrieval path. This article therefore analyzes Claude's citation logic on its own terms.
Google ranks documents. Claude assembles an answer. That difference sounds academic at first. Then you watch a brand win position one on a commercial keyword and get no mention when a buyer asks Claude.
The tasks people bring to Claude are not the tasks they bring to a search box. Anthropic's own usage research shows management-related work rising as a share of Claude.ai traffic. Investment memos and customer questions drive much of it. On the API side, business sales and outreach workflows at least doubled their share between late 2025 and February 2026. That category includes lead qualification research and customer data enrichment. The same research puts the average Claude.ai task above the US national wage average. Most of that work requires higher education.
Read that as an audience description rather than a vanity statistic. The person asking Claude which vendors to evaluate is often further along and closer to a real budget. Losing that answer costs more than losing a click.
Business users frequently reach Claude through a company account rather than a personal one. Forrester found that 61% of business buyers use AI tools their own organization provides. Many run private versions behind the firewall. Those deployments often carry internal context alongside public web retrieval.
You cannot see any of it. There is no referrer, no session, and no log entry on your side. The only signal available is what the model says when asked, which is exactly what prompt-level monitoring captures. This is also why positioning accuracy deserves attention equal to raw mention volume. A wrong description inside a procurement conversation costs a deal you never knew existed.
When Claude answers a question about your category, it draws on two distinct pools. One is what the model absorbed during training, which shapes how it describes your brand even with no live lookup. The other is live retrieval, where Claude fetches current pages to ground a specific claim. Those pools update on different schedules and respond to different signals.
Neither pool is a ranked list. There is no position three to climb toward. A source either enters the answer or it does not. The same prompt phrased two ways can pull entirely different sources. Consequently, the old habit of optimizing one page for one keyword produces unreliable results here.
Rank trackers built for blue links cannot see any of this. You need prompt-level monitoring that runs the actual questions your buyers ask. It should record whether your brand appeared, how it was described, and which domains got cited. Cognizo's Answer Engine Insights module handles that monitoring across supported platforms. Results break down by model, topic, prompt, and region, so the Claude picture stays separate from ChatGPT and Google AI Overviews.
Cadence matters as much as coverage. Answers shift as retrieval indexes refresh and third-party coverage changes. A monthly snapshot tells you almost nothing about cause and effect. Continuous tracking is what lets you connect a change you made to a change in the answer. Our guide to how Claude rank tracking works covers the mechanics in more depth.
Before you touch a single page, you need an accurate mental model of the pipeline. Most wasted effort in SEO for Claude goes into one stage while a different stage quietly blocks everything.
Anthropic documents three distinct bots, each with its own user agent and its own consequence when blocked. ClaudeBot collects web content that may contribute to model training. Claude-User retrieves pages when an individual asks Claude a question. Claude-SearchBot navigates the web to improve search result quality, analyzing content to make search responses more relevant and accurate.
The consequences are spelled out plainly in Anthropic's crawler documentation. Disabling Claude-SearchBot stops your content being indexed for search optimization. Anthropic warns that this may reduce your visibility and accuracy in user search results. Disabling Claude-User stops retrieval in response to a user query, which may reduce visibility for user-directed web search. All three honor robots.txt. Anthropic also notes that IP-based blocking is unreliable, since it interferes with reading robots.txt at all.
Search Engine Journal covered the update when it landed. The blanket "block AI crawlers" approach from 2024 no longer behaves the way teams assume. Blocking ClaudeBot stops training collection and does nothing about Claude-SearchBot or Claude-User. That is the single most important technical fact in this article.
Think of trained memory as the model's default description of you. It governs the answer when Claude responds without looking anything up, and it changes slowly. Live retrieval is the correction layer. It governs freshness, current pricing, recent launches, and anything a user pushes Claude to verify.
You need both. A brand with strong trained representation but blocked retrieval gets described from a stale snapshot. That is how outdated pricing reaches a buyer. A brand with open retrieval but thin trained representation appears only when the prompt happens to trigger a search. Neither situation is recoverable through content alone.
This failure is far more common than it should be. Search Engine Journal reported on an analysis of robots.txt files across 100 top news sites. It found that 79% block at least one AI training bot. More importantly, 71% also block at least one retrieval or live search bot. Within that sample, ClaudeBot was blocked by 69% of sites and Claude-Web by 66%.
Publishers had a reasonable argument for those decisions. Most B2B brands do not. If your robots.txt was copied from a 2024 template, it may currently remove you from the citation layer you are trying to win. Check it before you spend a dollar on content.
Any diagnosis has to separate two gates. Either the crawler cannot reach the page, or it reaches the page and still does not cite it. Most teams assume the second problem and have the first one. Work through the checks below in order.
Open your robots.txt and search for every Anthropic user agent by name, including the deprecated strings. Look for ClaudeBot, Claude-User, Claude-SearchBot, Claude-Web, and anthropic-ai. Then confirm what each directive actually does rather than what you think it does.
A reasonable default for a B2B brand that wants to be found looks like this. Allow Claude-SearchBot and Claude-User without restriction, since both feed live answers. Treat ClaudeBot as a separate decision about training data. Blocking it does not stop Claude describing you. It only stops your new material feeding future training. Repeat the audit on every subdomain, because robots.txt does not inherit across them.
Two practical traps show up repeatedly. Wildcard rules written as User-agent: * followed by broad disallows can catch Claude's bots without anyone intending it. Staging subdomains and documentation subdomains often carry blanket blocks that were never revisited after launch.
Retrieval bots do not wait around. If your key pages render their substance through client-side JavaScript, assume the fetched version may be close to empty. Server-side rendering or static generation for product pages, pricing pages, comparison pages, and documentation removes that risk entirely.
Response time matters for the same reason. Slow pages get fetched less reliably and contribute less often. Aggressive bot mitigation at the CDN layer creates a subtler version of the problem. Many WAF rules challenge unfamiliar user agents by default. Ask your infrastructure team to confirm that Claude-User and Claude-SearchBot receive a 200 response rather than a challenge page.
Keep your XML sitemap accurate and current, because discovery still depends on it. Beyond that, publish an llms.txt file. It gives a structured map of what your site covers and where the canonical answers live. Our guide to llms.txt explains the format and what belongs in it.
Schema markup earns its place here too. Organization, Product, FAQPage, and Article schema give retrieval systems unambiguous entity data instead of inference. That reduces the chance of Claude confusing you with a similarly named company. Such mix-ups are common in crowded B2B categories.
Cognizo's technical audits run these checks continuously rather than as a one-off project. Crawler access problems, rendering failures, and schema gaps get flagged as they appear. A crawl configuration that was correct in January can break silently after a CDN migration in March.
Once access is confirmed, the question becomes extractability. Claude has to pull a clean, self-contained statement out of your page and drop it into an answer. Pages written as flowing narrative make that hard, even when the underlying information is excellent.
Put the direct answer in the first 40 to 60 words under each heading, then expand. This is the single highest-leverage content change in SEO for Claude, and it costs nothing but discipline. A section that buries the point under three paragraphs of context gives a retrieval system nothing to quote.
On a pricing page, state the price and what it includes before explaining the packaging philosophy. On a comparison page, name the actual difference in the first sentence rather than building toward it.
Aim for sections of roughly 120 to 180 words. Each one should make sense if lifted out of the page. Avoid pronouns that reach back to an earlier paragraph. A section beginning with "this approach" loses its meaning once extracted.
Specificity beats adjectives every time. "Reduces onboarding time" is unusable. "Reduces onboarding from 14 days to 3" is a fact a model can cite. Numeric claims get extracted far more often than qualitative ones. Write dates, versions, figures, and named integrations into the copy rather than leaving them to a sales deck.
FAQ sections are among the most reliably extracted content on a page. That holds only when they answer questions people actually ask, rather than questions marketing wishes people asked. Pull the wording from sales call notes, support tickets, and your own prompt research.
Keep each answer between 60 and 120 words, write it to stand alone, and mark it up with FAQPage schema. A question that restates a heading wastes the slot. Aim instead at the edge cases and decision-stage concerns that the body of the page does not reach.
Give every page a named author with real credentials, a review date, and a clear organizational affiliation. Retrieval systems weigh those signals. Anonymous content signed by a company name carries less. For regulated and technical categories, this gap widens considerably.
Add a visible last-updated date and honor it. Stale dates on pages that clearly changed create the opposite signal from the one you want.
Business users ask Claude decision-stage questions constantly, and those prompts pull toward pages that answer them directly. Comparison pages, pricing pages, integration docs, and use-case pages outperform top-of-funnel explainers in commercial categories.
Cognizo's Content Optimization module works from your visibility data rather than from a generic checklist. It produces prioritized recommendations, then briefs, outlines, drafts, and FAQ blocks tied to the prompts where you are missing. When a prompt shows zero appearances for your brand, the brief it generates targets that specific gap.
Here is the part most teams underinvest in. Claude does not assemble its view of your category from your website. It assembles it from what the wider web says about your category, and your site is one source among many.
The pattern is consistent across AI platforms. Earned citations link to a third party rather than to your domain. They make up the larger share of what appears in answers. Review platforms, comparison articles, analyst write-ups, industry publications, and community threads all feed the pool.
That reframes the work. Instead of asking whether your site is optimized, ask which domains Claude already cites on your topic. Then work out what it takes to appear on them. Source mention rate answers the first half of that question. Cognizo reports it per prompt and per platform.
Software review sites carry weight in B2B because they are structured, updated, and independent. Keep your profiles complete, keep feature lists current, and keep review volume moving. A profile that has not been touched in 18 months describes a product that no longer exists.
Community discussion matters for a related reason. Long-form threads where practitioners compare options give retrieval systems exactly the kind of comparative language that answers comparative prompts. You cannot manufacture those threads honestly. You can make sure the people who would naturally join them understand your product.
In a guest article, a roundup, or a trade quote, the citable unit is the sentence containing your claim. Give writers specific, verifiable figures rather than marketing language. A publication that prints "cut reporting time by 60%" creates a citable fact. One that prints "streamlines reporting" creates nothing.
Cognizo's content toolbox covers PR, affiliate, social, and owned media as one connected surface. The same visibility data shows which off-site properties are worth pursuing. Our article on checking whether your brand appears in AI search covers the manual version.
Coverage is where most SEO for Claude programs are weakest, and it is rarely a content problem. It is a research problem. Teams track 20 prompts, see decent numbers, and conclude they are visible. The actual question space runs into the hundreds.
A single commercial intent generates a wide spread of phrasings. "Best invoicing software for agencies" sits alongside "invoicing tools that integrate with QuickBooks". So do "what should I use instead of spreadsheets for client invoicing" and dozens more. Each of those can return a different set of brands.
Your tracked set should reflect that spread rather than a tidy keyword list. Problem-phrased prompts deserve particular attention, since buyers frequently describe a symptom rather than name a category. Those prompts are also where incumbents are weakest, because nobody optimizes for them.
B2B purchases involve many stakeholders, and each one asks differently. A finance lead asks about total cost and contract terms. A security reviewer asks about SOC 2, data residency, and SSO. An end user asks whether it integrates with the tools they already have. An executive asks which vendor the market considers the leader.
Build prompt sets that cover each role explicitly. A brand can be well covered on practitioner prompts and invisible on security and procurement prompts. Enterprise deals stall there.
Cognizo's Prompt Volumes module is built on billions of real-world signals. It shows what buyers actually ask rather than what a keyword tool infers. It also generates prompt variations automatically and enriches them with CRM and support data. That surfaces the phrasings your customers used before they found you.
From there the sequence is mechanical. Rank prompts by commercial intent, check current appearance rate for each, and prioritize the high-intent prompts with zero appearances. More tracked prompts always produce a more accurate picture, so resist the temptation to trim the list for tidiness. Cognizo Growth tracks 150 prompts and Pro tracks 350, each across 5 selectable platforms. Enterprise covers all 10 supported engines, Claude included, with a custom prompt allowance.
You cannot manage this with a traffic dashboard. Most Claude answers carry no link at all. Referral data therefore captures a small fraction of the value.
Cognizo's framework uses six numbers, and each one answers a different question.
Treat Visibility Score as the headline number and the rest as diagnostics. A flat Visibility Score with rising earned citation share means the off-site work is landing but has not yet converted.
The cleanest mental model is the one from traditional search, shifted one layer up. Mentions and citations behave like impressions. AI-referred traffic behaves like clicks. It will always be smaller, because most answers do not link anywhere.
Teams that judge the channel on referral volume alone conclude it does not work. Teams that judge it on mention volume and pipeline quality reach the opposite conclusion. The same Forrester analysis lands on the second reading. It advises providers to shift from driving traffic through search engine optimization to driving visibility through answer engine optimization.
Since most Claude mentions carry no link, the realistic journey runs from mention to branded search to direct visit. UTM parameters never see the first two steps. Two fixes help. Add a "how did you hear about us" field with an explicit AI option. Then watch branded search volume and direct traffic.
Cognizo's AI Traffic Analytics connects crawler behavior to outcomes. It tracks ClaudeBot, GPTBot, OAI-SearchBot, and other major bots alongside referral sessions and conversions per platform. Seeing which bots fetched which pages, and when, closes a large part of the attribution gap. Our guide to measuring AI share of voice covers the benchmarking side in detail.
Absence is the obvious failure. Misrepresentation is the expensive one, because it looks like success on a visibility dashboard while actively costing you deals.
Claude may place you in an adjacent category that no longer fits, especially after a repositioning or a product expansion. A buyer who reads that description filters you out before ever visiting your site, and you never see the loss.
The fix is consistent category language everywhere the model can read it. That means homepage, product pages, review profiles, press coverage, and documentation. Inconsistency across those surfaces is what creates the ambiguity in the first place.
Trained representation lags reality. Suppose you changed packaging nine months ago and review profiles still list the old tiers. Claude now has several confident sources telling it something false. Audit every third-party property that states a price or feature limit. Update them on a schedule rather than when someone notices.
Sentiment moves when third-party coverage moves. A cluster of critical reviews or one widely read comparison article can shift how Claude characterizes you within weeks. Continuous sentiment tracking catches drift early, while it is still a content problem rather than a reputation problem. Our guide to tracking brand sentiment in AI-generated answers covers the workflow.
Spot checks miss this. Positioning errors cluster around specific prompt types, usually comparison prompts and category-definition prompts, and they vary by platform. Reading ten answers by hand tells you nothing reliable about the other two hundred.
Cognizo scores positioning accuracy across the full tracked prompt set. You can see which prompts produce a wrong category, which produce an outdated capability claim, and where platforms disagree. From there the remediation is specific: a set of pages to update, a set of third-party profiles to correct, and a short list of publications worth approaching with accurate information.
Five patterns account for most stalled programs.
Blocking retrieval bots by accident. Covered above, and still the most common single cause of total invisibility.
Tracking too few prompts. A 20-prompt set produces a comfortable number and a misleading one. Real coverage requires hundreds.
Treating all AI platforms as one surface. Retrieval differs per platform, so a ChatGPT-tuned strategy will not transfer. Analyze Claude on its own terms.
Measuring on referral traffic. Judged this way, the channel looks like a rounding error. Then sales starts hearing your name repeated back from AI research.
Optimizing only owned media. Your site is one input. The third-party layer usually carries more weight, and it takes longer to move, so it should start first.
Every step above is real work. Most teams lack the headcount to run it continuously across ten platforms. That is the gap Cognizo was built to close.
Autopilot is the content agent with auto-publishing, and it ships on every plan rather than sitting behind a premium tier. Agents handle market research, prompt planning, content production, and publishing. That turns SEO for Claude from a project into a running system.
A quarterly audit produces a backlog nobody clears. Here the loop closes continuously instead. A prompt gap appears, a brief gets generated, content ships, and visibility gets re-measured against the same prompt set.
Answer Engine Insights delivers real-time monitoring across supported platforms. It covers visibility, share of voice, citation share, source mention rate, sentiment, positioning accuracy, topical authority, and competitive benchmarking. Everything breaks down by model, topic, prompt, and region.
Content Optimization turns that data into prioritized recommendations, automated briefs and drafts, and technical audits for crawler readiness. Its toolbox spans PR, affiliate, social, and owned media.
AI Traffic Analytics connects crawler behavior to outcomes. It tracks Claude's bots alongside other major crawlers, plus referral traffic, sessions, and conversions per platform.
Prompt Volumes reveals what buyers actually ask AI. It runs on billions of real-world signals, with AI-powered prompt generation and CRM enrichment.
ChatGPT Ads pairs organic visibility with paid placement and surfaces competitor ad copy. It integrates the OpenAI Conversions API with Google Ads and Search Console.
Cognizo also captures rendered answers through UI scraping. The data therefore reflects what a real user sees, not API sampling alone. Our Claude rank tracker comparison goes deeper on that methodology question.
Growth is $499 per month. It includes 8 AI-optimized articles, Autopilot, 150 tracked prompts, and 5 selectable platforms. Pro is $999 per month, with 20 articles and 350 tracked prompts on the same feature set. Enterprise is custom and covers all 10 supported engines, Claude included. It adds tailored content volume, multiple brands and workspaces, a dedicated AI search strategist, and SSO, SLA, and enterprise controls.
Tracking Claude itself sits in the Enterprise platform set. Brands that treat Claude as a primary surface should scope that from the start. Growth and Pro cover the highest-volume platforms and the full content loop underneath.
MCP and API access ships on every paid tier, starting at Platform. Most of the category gates programmatic access behind an enterprise contract. That blocks data teams who want visibility metrics inside their own warehouse or agent workflows. Every tier also includes unlimited seats, regions, languages, all-time history, and export.
Sequence matters. The technical work unblocks everything else, so it goes first.
Audit robots.txt across every subdomain for all Anthropic user agents, including deprecated strings. Confirm that Claude-User and Claude-SearchBot receive clean responses rather than WAF challenges. Verify server-side rendering on pricing, comparison, and documentation pages. Publish or refresh llms.txt and validate your schema.
Build a prompt set that covers category, comparison, problem-phrased, and role-specific questions across your buying committee. Run it and record appearance rate, position, sentiment, and cited domains for each prompt. Separate the Claude results from every other platform, since mixing them hides exactly the divergence you need to see.
Take the highest-intent prompts with zero appearances and fix the pages that should answer them. Rewrite openings to lead with the answer. Break long sections apart, replace vague claims with figures, and add named author attribution. Ship comparison and pricing updates first, because those pages pull the most retrieval weight in commercial categories.
Update every third-party profile that states your pricing or feature set. Pursue placement on the two or three domains your source mention rate data shows Claude citing most in your category. Then re-run the full prompt set and compare against your day 12 baseline.
Thirty days moves technical access and owned content. Earned citations move on a longer clock, usually one to two quarters. Judge that work on leading indicators, not Visibility Score alone.
No, and this is the most misunderstood point in SEO for Claude. ClaudeBot only collects content that may contribute to model training. Blocking it excludes your future material from training datasets. It has no effect on Claude-User or Claude-SearchBot, which handle live retrieval. Claude can still describe your brand from what it already learned and from pages it fetches at query time. The reverse is more damaging. Blocking the retrieval bots while allowing the training bot removes you from live citations.
Technical fixes show up fastest. Once a retrieval bot regains access to blocked pages, changes can appear within days to a few weeks. Owned content rewrites typically take two to six weeks to register, depending on how often your pages get re-fetched. Earned citations run slowest, usually one to two quarters, because they depend on third-party publication cycles you do not control. Trained representation shifts on the longest timeline of all, tied to model release schedules rather than anything you do.
Yes, though the foundations overlap. Crawler access, extractable formatting, and strong third-party coverage help on every platform. The differences sit in retrieval behavior, index freshness, and source weighting. That is why the same prompt returns different brands on different surfaces. Claude also skews more heavily toward professional and enterprise use, so decision-stage and procurement-oriented prompts carry more relative weight there. Track each platform separately rather than assuming one number represents all of them.
Track mentions rather than clicks. Visibility Score, the percentage of tracked prompts where your brand appears, is the primary KPI. Support it with share of voice, citation share, source mention rate, sentiment, and positioning accuracy. On the traffic side, pair referral data with branded search volume and direct visits. The real journey runs from mention to branded search to direct visit. A "how did you hear about us" field with an explicit AI option captures buyers no analytics tool will attribute.
More than feels necessary. A tracked set of 20 prompts produces a number, not a picture. Real query space in a B2B category runs into the hundreds. Phrasing variations, comparison framings, problem-phrased questions, and role-specific angles all multiply it. Map the fanout around your top commercial intents first. Then expand into symptom-phrased prompts, where incumbents are weakest. Coverage gaps you never measured are the ones that cost the most.
Rarely. Content works only after two conditions are met. Claude's retrieval bots must reach your pages, and third-party sources must describe your category accurately. If access is blocked, no amount of publishing changes anything. If the wider web describes you wrongly, your own pages compete against a much larger volume of contrary signal. Sequence it as access first, owned content second, earned coverage third. Treat all three as ongoing work.
Growth is $499 per month and includes 8 AI-optimized articles, Autopilot, 150 tracked prompts, and 5 selectable platforms. Pro is $999 per month, with 20 articles and 350 tracked prompts. Enterprise is custom priced and covers all 10 supported engines, including Claude, plus multiple brands and workspaces, a dedicated AI search strategist, and SSO, SLA, and enterprise controls. Every plan includes daily tracking, unlimited seats, regions, languages, and competitors, plus MCP and API access.
The behavioral evidence is strong. Forrester reports that 94% of business buyers now use AI in their buying process, up from 89% a year earlier. Twice as many buyers name generative AI or conversational search their most meaningful information source. That puts it ahead of vendor websites, product experts, and sales. Forrester also notes that 61% of buyers use AI tools their own organization provides, often behind the firewall. A shortlist assembled inside those tools is formed before any vendor conversation happens.
Continuously. Answers change as retrieval indexes refresh, as third-party coverage shifts, and as competitors publish. A monthly check tells you something moved without telling you what caused it. That makes the data useless. Daily tracking against a fixed prompt set connects a specific change, such as unblocking a crawler, to a movement in Visibility Score. It also catches positioning errors and sentiment drift while they are still small enough to fix cheaply.
.svg)
.svg)
