For a while, "AI search" meant ChatGPT in most conversations. That is changing. Referral data from 2026 shows Claude taking a noticeably larger slice of AI-driven traffic, particularly among business and professional audiences. If your customers are developers, analysts, marketers or other knowledge workers, there is a decent chance some of them ask Claude before they ask a search engine.
By the iSuggest.ai Team ยท Updated for 2026
This guide explains what the data shows, how Anthropic's crawlers behave, and what you can do to make your website easier for Claude to use. As elsewhere, we avoid promises. Nobody can guarantee a recommendation. But the controllable parts are clear, and most of them also help with other assistants. For a general introduction, see what GEO is and why it matters.
What the traffic data says
Industry trackers report that referral traffic from Claude grew about 320% year over year, compared with 27% for ChatGPT's absolute traffic and 231% for Gemini, though all three started from very different bases (PPC Land). In B2B specifically, one analysis found ChatGPT's share of AI referrals falling from 89% to 63% while Claude's rose from 1.4% to 18.5% (SEO Sherpa).
A few cautions are in order. Percent growth from a small base looks dramatic, and these figures come from specific panels of sites, not from the whole web. Your own audience may skew differently. The practical reading is modest: Claude is a channel worth preparing for, especially if you sell to professionals, and a single-assistant strategy is increasingly fragile. For the ChatGPT side of the picture, see how to get cited by ChatGPT.
How Claude finds web content
Claude can answer from its training, and it can also use web search or fetch pages when it is set up to do so. Anthropic documents several crawlers with different purposes, commonly described as one for collecting training data, one that supports search indexing, and one used when a user asks Claude to fetch a specific page. Names and exact roles can change, so read Anthropic's current documentation before you edit any rules. Our post on the rise of AI crawlers gives the broader picture of who is knocking on your server.
What matters for you is the principle: a crawler you block cannot read your page, and a page it cannot read cannot be cited. If you block training crawlers on purpose, that is your call, but check that you are not blocking the ones that support live answers in the same stroke.
Step 1: Review your robots.txt and server rules
Open your robots.txt and look for rules that name Anthropic's user agents, rules that block all unknown bots, and wildcard disallows. Then check your firewall or CDN. Bot-protection services sometimes challenge or block unfamiliar crawlers by default, and that happens before robots.txt is even consulted. A crawler that gets a challenge page instead of your content reads a challenge page. Our guide to robots.txt syntax mistakes shows the patterns that cause accidental blocks.
One study cited in industry reporting suggests that more than 35% of the top 1,000 websites block at least one major AI crawler (as reported by LLM Pulse). Some of those are deliberate decisions. Many are probably inherited defaults. Decide, on purpose, what you want.
Step 2: Serve your content in plain HTML
Claude's fetching tools, like most AI crawlers, work best with content that is already in the HTML response. If your page builds its text with JavaScript after load, a fetch that does not run scripts sees little. View the source of your key pages and check that the headline, the main copy and the key facts appear in it. If they do not, consider server-side rendering for those pages. We walk through the test in what client-side rendering hides from AI crawlers.
Step 3: Write for a careful, professional reader
Claude is often used for research, analysis, writing and coding help, and its users tend to ask detailed, specific questions. Content that is precise, qualified and well sourced suits that style. A few habits help.
- State assumptions and limits. "This applies to teams under 50 people" is more useful and more quotable than a universal claim.
- Cite your sources. Link to the original research, with the figure and its framing intact. Pages that show their evidence are easier to trust and to reuse.
- Prefer specifics over adjectives. "Imports a 10,000-row CSV in under a minute" says more than "blazing fast", provided it is true and you can back it up.
- Define terms. A passage lifted out of context should still make sense.
Our guide to writing content AI models want to quote goes into these habits in more detail.
Step 4: Document your product like you expect to be asked about it
Developers and technical buyers frequently ask AI assistants things like "what does this tool do", "how does it compare" and "does it support X". If your documentation, pricing page and feature pages answer those directly, an assistant has accurate material to draw on. If the answers live only in a login-protected area or a PDF, they are invisible.
Make sure these pages exist as public, readable pages: what the product does, who it is for, what it costs, what it integrates with, and what its limits are. Be honest about limits. An assistant that has read your own clear statement of a limitation is less likely to guess, and less likely to guess wrong.
Step 5: Make your entity clear and consistent
An assistant has to be sure which company or product a page is about before it can recommend it. Use one consistent name, a clear description, and structured data that identifies the organization and links to official profiles. Entity clarity is one of the cheapest improvements and one of the most overlooked. Pair it with structured data as described in structured data for AI search engines, and keep the markup faithful to what is visible on the page.
Step 6: Earn mentions beyond your own site
Assistants do not only read your pages. They also reflect how the wider web talks about you. Industry analysis reports that brand mentions correlate with AI visibility about three times more strongly than backlinks do (0.664 versus 0.218, per Media Officers), and a Meltwater study found earned and news media accounted for 37.6% of language model citations versus just 0.2% for press releases (Meltwater). Treat the correlation carefully, but the direction is sensible: credible third parties describing you accurately is worth more than another self-published announcement. We explore the second finding in earned media versus press releases.
In practice, that means contributing real expertise to industry sites, keeping listings accurate, encouraging honest reviews, and being easy to describe. A one-sentence description of what you do, used consistently, helps everyone who writes about you.
Step 7: Keep things current
When a model uses live search, recency can matter. Show a real date, update pages when facts change, and review fast-moving topics on a schedule. Our recommended rhythm is in how often to re-audit your website.
Step 8: Audit, fix, and publish a snapshot
You should know what a machine sees before you decide what to change. Run a GEO audit on your most important URL. iSuggest.ai audits are free, and the report covers identity, structure, markup, links, images and performance. Read what we measure in an audit to turn the report into a short to-do list.
Once the page is in good shape, you can publish a clean, factual snapshot to the iSuggest.ai AI directory, which has a page for Claude. A snapshot is a crawlable record of what your page says at a point in time. Publishing spends credits, while audits do not, as explained in our credit system post. It improves findability. It is not a promise that Claude will cite you.
What to avoid
Avoid anything that promises guaranteed AI recommendations, anything that fakes reviews or mentions, and anything that hides text from readers but shows it to crawlers. Those tactics are not only unethical, they tend to backfire, because the systems that evaluate sources are good at spotting low-trust patterns. Avoid, too, the trap of writing for one assistant. The requirements for Claude, ChatGPT and Gemini overlap heavily: reachable, readable, specific, consistently described and trustworthy. Do the shared work first.
Common questions about Claude and your website
Do I need a separate strategy for Claude?
Mostly no. The fundamentals that make a page usable for Claude, access, readable HTML, clear answers, consistent identity and credible mentions, are the same ones that help with ChatGPT and Gemini. Where Claude differs is in audience: its users skew toward professional and technical work, so detailed, well-qualified content tends to fit. Build the shared foundation first and tune the tone second.
Should I allow Anthropic's crawlers?
That is a business decision. Some publishers block training crawlers because they do not want their content used to train models, and that is a legitimate position. Others allow everything because visibility matters more to them. The mistake is not choosing either way; the mistake is blocking by accident. Write down the policy you want, then check that your robots.txt and firewall match it.
How will I know if Claude is sending traffic?
Look at referral sources in your analytics for claude.ai and related domains, and compare before and after a round of improvements. Referral data from assistants is incomplete, because some answers are read without a click and some clicks arrive with no referrer. Use it as a trend line, not a precise count.
What if my content is in a PDF or behind a login?
Then assume it is invisible. Publish the essential facts, pricing, specifications, how-to steps and limits, as ordinary public HTML pages. PDFs can be fetched, but they are harder to parse cleanly and rarely carry structured data. Gated content can still exist, but pair it with a public summary that tells people and machines what is inside.
A worked example: a small software company
Imagine a small company selling scheduling software to clinics. This is an illustration, not a case study. Their homepage describes the product in marketing language, pricing is a PDF, and the integrations list is a JavaScript widget. A professional asking Claude "which scheduling tools integrate with our practice management system" would not get useful material from that site.
The fixes are plain. Replace the PDF with a pricing page in HTML. Render the integrations list on the server so it appears in the page source. Rewrite the homepage opening to say who the product is for and what it does in two sentences. Add Organization markup with official profiles. Publish an honest "who this is not for" paragraph. None of that is clever, and all of it makes the site easier for any assistant to describe accurately.
A one-week plan
Day one, read your robots.txt and your firewall settings. Day two, view source on your top five pages and confirm the content is there. Day three, rewrite the opening of each so it answers the main question. Day four, add or fix Organization and FAQ markup. Day five, write a plain, honest description of your product and put it on a public page. Day six, run an audit and record the score. Day seven, plan one piece of outside coverage you could earn this quarter. Then repeat the audit in a month.
If you would like to start with the audit, see how it works, then run it on your own URL and use the report to decide where to begin. When you are ready to publish a snapshot, sign in at your account page.