When Clauderecommends,why not you?
More than half the time, Claude answers from memory, not search. We mapped how it decides, and how to earn a place on the shortlist.
≈90,000 answers analysed · 15+ industries
of Claude answers come from memory, with no web search at all.
brands named in a typical recommendation. Prescribed, not browsable.
of prompts trigger a live search. On those, other people’s pages win.
Based on approximately 90,000 sources across 15 or more industries. Claude’s search behaviour is inferred from the network requests the app makes, observed through the interface rather than the API.
A Claude recommendation is a short list. Ask it for the best option in your category and it will name three or four brands, prescribed to the reader, not offered as ten blue links to browse. That scarcity is the whole game. And more than half the time, the names in that list were chosen before Claude read a single web page.
This is what makes Claude different from Google, and from ChatGPT. It leans on what it already knows, its parametric memory, and treats a web search as a way to confirm a view it already holds. For considered, reputation-led categories, the answer is effectively decided in memory. Search optimisation only applies to the half of questions where Claude actually searches. For the other half, the work is reputational, and it moves on training-cycle time, not publish-this-week time.
The rest of this study lays out the mechanics: how the balance shifts by industry and by question, what shapes the memory, where Claude looks when it does search, and how to measure whether it knows you at all.
Part one · The diagnosis
Claude relies on memory far more than ChatGPT.
Same prompt set, two very different engines. The habits that win on ChatGPT are not the habits that win here.
Behaviour by industry
Some categories run almost entirely on memory.
Considered, reputation-led categories such as software and health skew to memory. Live and local questions skew to search.
Share of answers drawn from memory, with no web search, by category.
Behaviour by question
The commercial questions are the memory questions.
Advice, comparisons and “is it any good” run on memory. Shopping and local questions are what trigger a search.
A head-to-head comparison is answered from memory almost nine times in ten. The shortlist forms from reputation, not from the comparison pages you built.
Even when it searches
The funnel narrows twice.
On a searched prompt Claude opens about seven pages and quotes about three. It is far more reluctant than ChatGPT to repeat a page it just found.
Part two · What shapes the memory
The shortlist forms before the search starts.
Take a thousand unbranded buying questions and read the internal queries the model writes for itself before it reads a single page. One in six already name a specific brand that nobody put in the question. Those names were not retrieved. They were remembered.
of those internally named brands sat in the model’s own top-ten recalled list for the category, against a 2.6% random baseline.
internal searches named a brand on their own, unprompted. The category leader was chosen in memory, then confirmed by search.
Who the model misses
New and changing brands are the easiest to leave off the list.
New entrants
Launched since the model last learned? You are light in its memory. Not permanent, but the gap to close first.
Rebrands and renames
The model may still hold the old you. Redirects, entities and consistent naming tell it who you are now.
Challengers to an incumbent
The default answer is the name everyone knows. The work is giving the model reasons to say yours too.
Fast-movers
A new product, a pivot, a reputation you have already fixed. The memory has not caught up yet.
How memory is built
It runs on your reputation as the model absorbed it.
You are not editing a page and watching a rank move. You are changing what the wider internet says about you, so the next time the model learns, it learns something better.
Be described the same way everywhere
A “platform” here and a “tool” there blurs the picture. One clear category claim, repeated across trusted sources, settles into the model as a fact about you.
Be a custodian of your own brand
Earned media is downstream of what you say about yourself. Surface your positioning, your features and your proof so publications can learn it and repeat it. That is the citation loop.
A result that moved
In one controlled test, restructuring a section into question-shaped pages produced the first AI citations within four days, and about 2.6x the citations three weeks later. It can change.
Where to get written about
The sources Claude trusts, by category.
Get onto the pages Claude already reads for your category. Two names win across all of them: Forbes and TechRadar.
Software / SaaS
Get into
G2, Capterra, GetApp, plus Forbes and vendor round-ups
The move
Own your category on the review aggregators and the “best X software” listicles.
Consumer tech
Get into
RTINGS, TechRadar, Tom’s Guide, Consumer Reports
The move
Get the product independently tested and reviewed by the specialist testers.
Finance
Get into
NerdWallet, Forbes, CNBC-class editorial
The move
Earn comparison and “best of” coverage in finance editorial.
Home & garden
Get into
Category review sites and specialist testers
The move
Get into the niche testing sites, where Claude quotes most generously.
Any category
Get into
Forbes and TechRadar
The move
The two dual-winners. Worth pursuing whatever your category.
Part three · The searched half
When Claude searches, it is reading Brave, not Google.
Behavioural inference from what Claude fetches. The ranking work poured into Google has been aimed at the wrong index.
in Brave’s top ten
in Google’s top ten
Share of the pages Claude fetched that sit in each index’s top ten for the same query. When the two lists differ, Claude follows Brave by roughly six to one.
Inside the top ten, position does not matter
Claude grabs Brave’s top ten as a set, then reorders by relevance to the prompt. Quote rate is roughly flat whether a page sat at rank one or rank ten. The target is not “rank number one”. It is “be anywhere in Brave’s top ten for your category”. A lower bar, and a clearer one.
The one move
Half the answers never trigger a search.
The one move
Half the answers never trigger a search.
If you only did one thing
Write less about yourself.
Get more written about you.
Memory is built from what others say about you. And when Claude does search, it trusts other people’s pages over yours by two to one.
Owned versus earned
On Claude, a page on your own domain is worth about half what a mention on a site it trusts is worth. That reweights the whole plan: less polish our pages, more get onto theirs.
ChatGPT uses your own site
Claude uses your own site
It is an earned-media engine, not an owned-media one.
Before anything else
If it can’t be read, it can’t be cited.
A blocked page does not drop in rank. It simply never appears, in any answer, ever. And nothing warns you.
The crawler is blocked
ClaudeBot respects robots.txt, so a stray rule shuts it out. Worse, Cloudflare or Akamai can block AI bots by default on a prefetch rule, with nothing in robots.txt to warn you.
The page needs JavaScript
AI crawlers do not run JavaScript; Google is the exception, not the rule. If content only appears after a script runs, Claude sees a blank page. Serve the words in the HTML, render server-side.
The silent-block test
Ask Claude to fetch the H1 from your page’s source. If it cannot, you have been silently blocked. Anthropic’s crawler converts at roughly 2.4% crawl to referral: crawling is necessary, not sufficient.
What to de-prioritise
The old reflexes, and what they do on Claude.
None of these are “never do them”. They are “do not let them be your Claude plan”.
Field notes
Six things we learned watching Claude work.
The details behind the headline numbers, from observing how the model behaves prompt by prompt.
Memory is a hypothesis it then tries to confirm
Claude forms a view from training data first, then, when it does search, uses the pages to back up or adjust that view. Search is corroboration, not discovery. Your reputation sets the starting position.
The shortlist is written before the search runs
Across 1,000 unbranded buying questions, the internal queries Claude wrote for itself already named specific brands one time in six, against a 2.6% random baseline. The category leaders were decided in memory.
Query fan-out is far narrower than ChatGPT
When Claude does search, it issues far fewer sub-queries. For local intent it appends the country (“best plumber in Manchester UK”) because local facts change. Less breadth, more weight on what it already believes.
It reads more than it quotes, on purpose
On a searched prompt Claude opens about seven pages and quotes about three. On health it reads three and quotes fewer than one. It is selective about who it repeats, which raises the bar on source quality.
Silent blocking is the quiet killer
ClaudeBot respects robots.txt, and a CDN can block AI bots on a prefetch rule with nothing to warn you. Test it: ask Claude to fetch your H1 from the source. If it cannot, you are invisible, and nothing drops in rank to signal it.
The audience skews to high-trust buyers
Claude’s user base leans toward finance, technology, engineering, and clinical professionals. In the categories those people buy for, being the name Claude already trusts is worth more than any single page.
Part four · The test
Does Claude name you, unprompted, when nobody asked it to?
The instinct is to ask “does Claude cite me”. Half the time it is not citing anyone, it is recalling. So measure recall, not citations.
- 01
Read the memory
Ask Claude to name the top brands in your category. That list is its memory out loud. Find your row, or notice you have none.
- 02
Watch it leak
Ask the real buying question, framed the way a customer would, with no brand in it. See whether the names from step one walk into the answer.
- 03
Count, do not eyeball
Roughly 98.7% of brand appearances change from one run to the next. Run the question about forty times and count how often your name returns.
- 04
Weight it 90 / 10
Track mostly unbranded prompts (around 90%), a few branded (around 10%). Branded prompts always surface you; unbranded prompts tell you the truth.
Intellectual honesty
What we can say, and what we can’t.
What we can say
- The shortlist forms in memory, before any search.
- You can measure whether your brand is on it.
- You can shape what the model learns about you over time.
What we can’t
- That being on the list guarantees you win the answer.
- That any one tactic reliably puts you there.
- That you can move it overnight.
On Google you earn a ranking. On Claude you earn a reputation.
We run the diagnostic on your brand and tell you exactly where you stand in the model’s memory, then build the plan to move it. That plan is the engagement, not a download.