How to Get Cited by Claude: 6 Steps

Claude only cites when it searches, and it does not search for everything. The share of your commercial prompts that Claude answers with no search and no sources is what I call your Silent Answer rate, and it is the first thing to measure, because on those prompts your brand is either in the model’s memory or it is nowhere. For the prompts where Claude does search, three things separate it from every other engine: it runs on an index that is almost certainly not Google’s, it cites a median of about 3.6 domains per answer, and it barely touches forum content. Six steps follow. Budget two weeks. Claude passed Perplexity in referral sessions in March 2026 and now sits second behind ChatGPT for B2B audiences, so if you sell to businesses this is no longer optional. If you sell aircond servicing in Cheras, it still is.

What you’ll need before starting

Table 1. Prerequisites, with real costs and time.

Realistic effort: about two working weeks, most of it in steps 4 and 5.

Why this works

Quick Answer: Claude answers from what it already knows unless the question pushes it to search. When it does search, Anthropic’s own documentation confirms it retrieves current web content and returns citations with source URLs. The retrieval runs through third-party search infrastructure rather than Google, which is why your Google rankings transfer poorly and why a page can be ranked well and still be invisible here.

Three mechanisms matter, and I want to grade the evidence for each rather than present them all as equally solid.

The crawlers are documented, and this is the firm ground. Anthropic updated its crawler documentation around 20 February 2026 to formally separate three bots. Each has its own user-agent string, each can be allowed or blocked independently, and blocking one does not block the others.

Table 2. Anthropic’s three crawlers.

That robots.txt position is worth pausing on, because it differs from the competition. Perplexity states its user-initiated fetcher generally does not honour robots.txt, and OpenAI warns that robots.txt rules may not apply to ChatGPT-User. Anthropic says all three of its bots comply. Practically, that means your robots.txt file is a real control here and not a suggestion, in both directions.

The search index is strongly evidenced but not officially confirmed. Anthropic added Brave Search to its published subprocessor list in March 2025. Simon Willison found a BraveSearchParams parameter inside Claude’s web search tool definition within days of launch, and reported matching citations between Claude and Brave. Brave wrote in February 2026 that its Search API supplies most of the top-ten LLMs and that for some of them it is the only index behind their AI answers. Anthropic has also since listed a second web-search subprocessor, TurboPuffer, added in May 2026.

Here is where I break with most guides on this topic. The figure everyone repeats is an 86.7% overlap between Claude’s citations and Brave’s top results. That number is 13 out of 15 results. Fifteen. I am not going to build your strategy on a sample that size, and neither should anyone selling you a Claude package.

Table 3. What we actually know, by evidence tier.

Claude’s search behaviour differs measurably from ChatGPT’s. In a 2026 comparison by Josh Blyskal, Claude and ChatGPT shared only 8% of citation domains on average. Claude cited listicles in 36.4% of the classified set against 19.7% for ChatGPT. Forums and user-generated content made up 15.8% of ChatGPT’s citations and 0.9% of Claude’s, a gap of more than seventeen times. And Claude appended a year to 94% of its fan-out queries, against 17% for ChatGPT.

Key takeaway: the crawler rules are solid ground, the index is a strong inference, and the behavioural differences are the most actionable part.

Step 1: Measure your Silent Answer rate

Quick Answer: Run your 20 commercial prompts in Claude and record two things for each: whether Claude searched at all, and whether your brand appeared. The share it answered without searching is your Silent Answer rate, and those prompts cannot be won with content optimisation.

You can see this directly in the interface. When Claude searches it shows the query it ran and the sources it used. When it does not, you get an answer with no citations, drawn from what the model already holds.

Split your 20 prompts into three buckets: searched and cited you, searched and cited someone else, answered silently. The three buckets need three different responses, and most agencies sell one.

What “done” looks like: a simple sheet with 20 rows and those three columns, dated. Re-run it monthly with the same prompts.

Common mistake: treating a silent answer as a content failure. It is an entity problem. No page you write this quarter changes what the model already believes about your category.

Step 2: Set your robots.txt deliberately, one bot at a time

Quick Answer: Allow Claude-User and Claude-SearchBot if you want citations, and decide on ClaudeBot separately, because blocking training does not block search.

# Search and citations
User-agent: Claude-SearchBot
Allow: /

User-agent: Claude-User
Allow: /

# Training: your call, independent of the above
User-agent: ClaudeBot
Allow: /

The training decision is a genuine business judgement and I am not going to pretend it is obvious. Blocking ClaudeBot keeps your content out of future training runs, which some publishers want. It also means the model is less likely to know who you are when it answers silently, which is the exact scenario step 1 is measuring. Those two goals conflict. Pick knowingly.

Common mistake: the blanket Disallow: / inherited from a staging config, or a copied robots.txt that blocks the deprecated anthropic-ai string and nothing current.

Step 3: Get indexed somewhere other than Google

Quick Answer: Because Claude’s web retrieval runs through third-party search infrastructure rather than Google, your Google position does not carry over. Check whether the likely underlying index has your pages at all.

Search your money queries on Brave directly and see what comes back. If your pages are missing entirely, that is a discovery problem no amount of on-page work fixes. Brave provides a URL submission page for requesting a re-fetch, useful after a new page goes live or a meaningful update.

Two honest limits on this advice. A re-fetch request only helps discovery, it does not improve the page or force a citation. And the “ignore Google entirely” line that circulates in this space gets crawl access backwards, since the technical hygiene that makes a page crawlable serves every index at once.

Common mistake: assuming a strong Google ranking means you are findable. Different index, different result set.

Step 4: Put the year in the title, and keep it true

Quick Answer: Claude appended a year to 94% of its fan-out queries in one 2026 comparison, against 17% for ChatGPT. Content that carries a current date in the title and the visible page is structurally advantaged in that retrieval.

This is the cheapest win in the whole guide and it comes with a condition. Update the content when you update the year. A “2026” title on a page whose figures are from 2024 is the kind of thing that gets a source dropped once a model reads it, and it is dishonest besides.

Pair it with a visible last-updated date and a dated provenance line on every data claim: method, sample size, collection window. That line is what makes a figure quotable.

What “done” looks like: every priority page carries a current year in the title tag, an on-page last-updated date, and at least one figure with its collection window stated.

Step 5: Structure as a list where a list is honest

Quick Answer: Claude cited listicle-format content in 36.4% of a classified citation set, nearly double ChatGPT’s 19.7%. Pages that gather options under a current, query-shaped title match what this retrieval surfaces.

For a Malaysian professional services firm that means the comparison page, the options roundup, and the “which is right for your situation” page are worth more here than the long essay. Give each item a heading, a one-line verdict, and a specific figure.

The condition is the same as step 4. If your topic is not a list, do not fake one. A page that implies comparison without providing it is a liability under Google’s spam policies and it reads as thin to a model too.

Common mistake: converting a genuinely explanatory page into a list and losing the explanation that made it worth citing.

Step 6: Do not reuse your ChatGPT forum strategy

Quick Answer: Forums and user-generated content accounted for 0.9% of Claude’s citations against 15.8% of ChatGPT’s in the same comparison. If you built a Reddit programme for ChatGPT visibility, it is doing close to nothing for Claude.

This is the finding I would act on fastest, because it is a reallocation rather than an addition. The hours going into community seeding for Claude visibility are better spent on the listicle-shaped pages in step 5 and the dated data in step 4.

Keep the community work if it earns you ChatGPT and Perplexity citations, and both platforms lean on it heavily. Just stop counting it as Claude work.

Common mistake: running one “AI visibility” programme across all engines and assuming the tactics transfer. Claude and ChatGPT shared only 8% of citation domains in that study. They are close to different channels.

How to know it’s working

Table 4. Leading and lagging indicators.

Source: Ai Mode reporting practice, August 2026. Timelines are our estimates from client work, not measured benchmarks. Referral figures undercount, since not every session passes a referrer.

Claude cited a median of about 3.6 unique domains per answer in a 1,200-prompt study run between April and May 2026, against 6.4 for Perplexity and 3.1 for ChatGPT. Fewer seats than Perplexity’s Citation Slot Surplus, more than Gemini’s. Set expectations accordingly.

Troubleshooting

Claude answers your category prompts without searching at all. High Silent Answer rate. This is entity work, not content work: consistent business details everywhere, third-party mentions, and enough independent corroboration that the model has a settled view. Expect six to twelve months.

Claude searches but cites competitors every time. Check Brave for the same query. If competitors appear there and you do not, the problem is discovery in that index, not your copy.

Claude-SearchBot never appears in your logs. Check for a blanket disallow, a WAF rule, or a deprecated user-agent string in robots.txt. Anthropic states it will not bypass CAPTCHAs, so an aggressive bot challenge will simply stop it.

You were cited last month and not this month. Answers regenerate and rankings in the underlying index shift. Re-run the prompt several times across a week before treating it as a loss.

Your ChatGPT visibility is fine and Claude is zero. Expected, given an 8% domain overlap between the two. Work the Claude-specific steps rather than assuming something broke.

Should Malaysian businesses prioritise Claude?

Quick Answer: If you sell to businesses, yes, and sooner than most local agencies are telling you. Claude overtook Perplexity in monthly referral sessions in March 2026 and reached roughly 18.5% of measurable B2B AI referrals in one March–April 2026 brand panel, second only to ChatGPT at 62.6%. If you sell to consumers locally, no. Claude’s audience skews professional, and your buyers are still in the Map Pack.

Table 5. Where Claude sits among AI referrers.

Source: referral shares from Goodie’s AI search traffic report, brand-averaged across a GA4 panel for March–April 2026; citations per answer from Attrifast’s 1,200-prompt study across 12 verticals, 51,723 citation events logged between 12 April and 14 May 2026. Retrieved 19 August 2026. The two columns come from different studies with different samples and are placed side by side for orientation, not as a single ranking. Referral panels skew B2B and Western, so treat the shares as directional for Malaysia rather than local measurement.

Two caveats I would want stated if I were the one buying. Standalone assistant referrals remain roughly one percent of total website traffic across the large published benchmarks, so this is a growing channel and not yet a large one. And a separate analysis found ChatGPT sending the overwhelming majority of trackable standalone AI referral traffic, which tells you the B2B panel above is measuring a specific slice rather than the whole web.

The honest summary: Claude is the fastest-moving of the four and the one most Malaysian agencies have not looked at. That is an opening, not an emergency.

Conclusion

Start with step 1, because it tells you which of the other five steps are even relevant. A high Silent Answer rate means your problem sits in entity consensus and no amount of page work will move it this quarter. A low one means the citations are winnable with dated, list-shaped, properly indexed content.

The rest is unglamorous. Set three robots.txt rules deliberately instead of inheriting them. Check whether a non-Google index has your pages. Put the year in the title and make it true. And stop counting your Reddit work as Claude work, because at 0.9% of citations it is not.

One last thing, and it is the reason I wrote the evidence tiers into this piece rather than a tidy list of rules. Most of what circulates about optimising for Claude traces back to a fifteen-result test from launch week. Build on the documented crawler behaviour, treat the index inference as an inference, and re-check it quarterly.

Frequently Asked Questions

How long does it take to get cited by Claude? For prompts where Claude already searches, expect movement in 4 to 8 weeks once your pages are indexed in the underlying search index and structured to be liftable. For prompts Claude answers silently with no search, the timeline is much longer, typically 6 to 12 months, because that requires shifting what the model knows rather than what it can retrieve.

What is the Silent Answer rate? Your Silent Answer rate is the share of your commercial prompts that Claude answers without running a web search, producing no citations at all. On those prompts your brand is either already in the model’s knowledge or absent entirely. It is the first thing to measure, because content optimisation cannot fix a silent answer.

Does Claude use Google’s search index? Almost certainly not. Anthropic added Brave Search to its published subprocessor list in March 2025, and Claude’s web search tool was found to carry a BraveSearchParams parameter. Anthropic has never officially named a provider, so treat this as strongly evidenced rather than confirmed. The practical implication holds either way: your Google rankings do not transfer automatically.

Should I block ClaudeBot? That depends on a real trade-off. Blocking ClaudeBot keeps your content out of training runs but makes it less likely the model knows your brand when it answers without searching. Blocking it does not block Claude-SearchBot or Claude-User, so search visibility is a separate decision. Set all three deliberately.

Is Claude worth optimising for in Malaysia? For B2B and professional services, yes. Claude passed Perplexity in referral sessions in March 2026 and reached about 18.5% of measurable B2B AI referrals in one 2026 panel. For local consumer services, no. Claude’s user base skews professional, and Malaysian local-intent searches still resolve in Google’s Map Pack.

Do my ChatGPT tactics work for Claude? Largely not. One 2026 comparison found the two engines shared only 8% of citation domains. Claude cited listicles at 36.4% against ChatGPT’s 19.7%, while forum content made up 0.9% of Claude’s citations against 15.8% of ChatGPT’s. Keep the community work for ChatGPT, and build list-shaped, dated pages for Claude.