If your team reports on "Claude citation share," you’re about to get a reality check because the number you see today measures something completely different than it did just three weeks ago.
That’s because Claude changed its default search index and didn’t tell anyone …
We found this because of a friend. Fenil, the founder of Open Benchmarks, noticed their site stopped getting cited in Claude. That sent our team at Writesonic into the data to date the change and test it. We ran the same 50 prompts through both of Claude's search modes (standard and extended) on three models (Opus, Sonnet, and Haiku) and compared every cited domain.
In short, two things changed:
- Claude's default search now runs on a new index. It isn't Brave, the search engine Claude used to rely on, and it isn't Parallel either.
- Claude now has two search modes. Standard mode is the default and uses the new index. Extended mode pulls from a much wider set of sites and looks closer to the old Brave-based search.
The two modes share just 10% of cited domains. And top citation sources like Wikipedia and Forbes rarely show up in standard mode.
If your brand has relied on these sources to show up in Claude, you could be losing ground in AI answers. Below, I'll walk through what changed, what the new index actually cites, and what I'd do about it if Claude matters to your AI search visibility.
The Switch Happened Quietly On October 5th
We didn't stumble onto this from one weird answer. We looked back at 648 search calls across 11 months of history and found a clean break.
The last “old-style” search call we saw was on September 27th. By October 5th, every call looked different (very spooky stuff, indeed).
The tell was a hidden setting in Claude's search calls. Before October 5th, Claude's web search tool never sent a "mode" field. Since October 5th, every call has one, set to either "standard" or "extended."
The timing lines up with the Haiku 5.5 release, but Anthropic's October 7th release notes don't mention it.
Two caveats up front:
- This is one account, so October 5th is when the switch happened for us. It isn't a confirmed global rollout date.
- We also measured sources. Not answer quality. I'm not sure if Claude's answers got worse or better, because we didn't test that.
What we can say is that the change applies across Claude's current models. Opus 5.5, Sonnet 5.5, and Haiku 5.5 all run the new index by default.
Claude Is Now Two Different Search Modes
Claude has two search modes.
- There's standard mode, the default most people get;
- And an extended mode that searches more broadly.
You'd expect the two to cite mostly the same sites, with extended adding a few more. They don't.
Across our 50 prompts in 16 categories, the two modes shared just 10% of cited domains. The median was 9%, and three prompts had zero overlap at all.
That means a brand can be visible in one mode of Claude and invisible in the other.
The footprint is different too:
- Standard mode ran 2.6 searches per prompt and cited 16 unique domains per answer.
- Extended mode ran 4.0 searches and cited 29 domains.
Claude's smaller models search even less. Sonnet 5.5 averaged 1.2 searches per prompt and Haiku 5.5 averaged 1.1. Yet both cited the same obscure sites as Opus. A lighter search doesn't mean a different set of sources.
Key takeaway: Standard mode pulls from about half as many sites as extended mode does for the same question. That's a lot of sources your brand might be counting on that Claude never cites.
We also checked whether standard mode is still Brave search under the hood. We ran the same prompts against Brave's own API, and the overlap with Claude's standard mode came back at 13.3% (extended mode jumps to 38.4% overlap). Whatever standard mode is running on now, Brave isn't the primary source anymore.
We ran the same test against Parallel, a search provider Anthropic lists as a web search subprocessor, and overlap came back at 11.3% on standard mode, and 18.6% on extended mode.
Almost none of the obscure sites Claude's standard mode leans on showed up in either one. So the new default is either an index Anthropic built itself or a provider nobody has identified yet.
Google Authority Doesn't Transfer
Ranking in Google and earning mentions on big, trusted sites has been one of the most reliable ways to show up in AI answers. It still works, and it was never the only factor. But in Claude's standard mode, it carries a lot less weight.
Wikipedia showed up in 14 of 50 extended mode answers. In standard mode, it showed up once. That's 14 times less often (on one of the most cited sites on the internet).
Forbes followed the same pattern, with 12 citations in extended mode and one in standard mode. TechCrunch was nearly absent too.
So who is Claude citing?
We opened every top domain that appeared only in standard-mode answers and documented each one. Six of the top 10 were directories, content farms, or sites that were technically broken when we visited.
| Domain | Domain Rating (DR) | Standard-mode citations (of 50 prompts) | What is it? |
|---|---|---|---|
| toolradar.com | 57 | 9 | A software directory that sells SEO, backlink, and paid-listing services to the vendors it covers. |
| dupple.com | 54 | 6 | The media company that owns Toolradar, so two of the top 10 belong to the same company. |
| mayfairfarms.com | 20 | 6 | Branded as a wedding and event-planning guide, but publishes reviews of ball python enclosures and gaming PCs. |
| csr.hdsupply.com | 57 | 6 | A subdomain of industrial supplier HD Supply with an expired SSL certificate, so the page Claude cites won't load. |
| subger.com | 9 | 4 | Roughly 12,000 subscription pages that read as templated, with prices that don't match across its own pages. |
| temperstack.com | 30 | 4 | Describes itself as a software review site, but its homepage currently returns a 404. |
One small pricing site, costbench.com (DR 61), stood out more than the rest. It's a software pricing database with an AI "buying advisor" built on top. It discloses affiliate links. However, the site list does not list a founder, a company name, or a founding date. It showed up in 11 of 50 standard-mode answers (22%), across unrelated categories like SaaS tools, healthcare software, legal tooling, and fitness equipment.
To be clear, I'm not saying those sites are scams. I'm saying the new index rewards a different kind of source than Google does. Long-tail directories and review sites carry weight. Domain strength carries less.
The clearest example is a closed-beta startup that launched this year. It got cited in 6 of 50 standard-mode answers (12%). A product most people can't even sign up for yet is showing up more often than Wikipedia.
The #1 brand in ChatGPT changes 52% of the time.
Get the report and find out why.
Claude Switches To Extended Mode For Shopping Questions
On shopping prompts, Claude sometimes escalates to extended mode on its own (without being asked). That happened on 5 of 50 prompts with Opus and 7 of 50 with Haiku.
Our data shows when this happens, not why. It's most likely how Anthropic set up search for buying questions. Adding "(extended mode)" to a prompt roughly doubles the sources Claude reads. It also brings sites like Wikipedia and Forbes back into the answer. There's no toggle for this in Claude's interface. Typing "(extended mode)" is the only way a user can force it.
That's helpful for anyone using Claude for research. But buyers aren’t going to type “(extended mode)” into Claude (or even know that they can do this), so standard mode, the default, is the version you need to show up in.
Your Claude Numbers Broke On October 5th
This is important if you’re tracking your AI search presence in a visibility platform. Every "Claude citation share" number from before October 5th measured a different system. Your trend line didn't dip or spike; it reset.
That's the danger of blended AI visibility scores. Combine ChatGPT, Gemini, Perplexity, and Claude into one number, and a reset like this one is hidden in the average. You'd see a small dip and move on, while your brand lost citations and share of voice in Claude's standard-mode answers.
This isn't the first time an engine has shifted its search behavior. We saw something similar when GPT-6 Astra started searching less and reading brand pages directly. Engines change how they find sources, and rarely do they let you know.
What Should You Do About It Now?
You need to find out which sources Claude trusts in your space. Then prioritize getting added there.
Here's where I'd start …
Track your visibility by AI engine
Stop reporting one blended AI visibility number. A drop in Claude can hide inside an average across ChatGPT, Gemini, Perplexity, and any other engines you are monitoring. That's why you need an AI visibility platform that shows your visibility and citations for each AI engine you monitor. When one engine moves on its own, you'll see it.
For Claude, go one level deeper and track it by model and by topic. Shopping questions switch to extended mode more often than SaaS questions do, so the two can look like different search engines.
You can see this in Writesonic by filtering by platform:
Check your Claude numbers around October 5th
If your AI visibility platform checks prompts daily, the data is already there. Set the date range to cover the past few weeks and look at Claude on its own. Did your visibility or citation share drop around October 5th?
That's when we saw the switch, but the date isn't official, so check a few weeks on either side too.
Find the sources Claude trusts in your category
Look at which directories, review sites, comparison pages, and other third-part websites show up in those answers. Some might be low quality. Some will be easy places to get listed or corrected. Some will require more effort and a dedicated outreach strategy. Prioritize the ones that are cited most often and appear across several of your prompts.
Make sure Claude's crawlers can reach your site
Anthropic runs three crawlers:
- ClaudeBot
- Claude-User
- Claude-SearchBot
Claude-SearchBot is the one that indexes pages for search, so blocking it can prevent you from showing in results. Check that your robots.txt allows all three. Then check the bot settings in Cloudflare or any other CDN you use. Many sites block AI crawlers there without knowing it.
Check whether your site is in Common Crawl
Search the Common Crawl index for your domain, and make sure your robots.txt allows its crawler, CCBot. We don't know exactly what Claude's new index is built on. But costbench.com, the site standard mode cited most, is in Common Crawl. Open Benchmarks, the site that tipped us off, isn't.
Publish clear pricing and comparison pages
A lot of what standard mode cites is pricing pages. When we asked for the best CRM, Claude searched for HubSpot's pricing and cited five other companies' HubSpot pricing pages. None from HubSpot itself. Clear, current pricing and comparison pages on your own site give Claude an accurate source to cite.
Watch for the next shift
This won't be the last unannounced update. Set up alerts when your AI visibility or citation share for your monitored prompts move suddenly.
In Writesonic, you can do this by setting up an alert and configuring the parameters of the alert (e.g., notify me when Citation Share is less than 15% compared to the last 7 days). You can also limit this by platform (e.g., Claude, ChatGPT, Gemini, etc.)
Claude Visibility Is Its Own “Can Of Worms”
Claude didn't just tweak its search. It moved to a different set of sources. The authority that carried you in Google doesn't carry over.
A closed-beta startup is out-citing Wikipedia in Claude's standard mode. That alone should tell you how much this new search index has shifted results. That doesn’t mean you should panic. It means you can't treat "AI search" as one channel with one visibility score. Track each AI engine on its own, know which sources it cites, and check often.
The brands that spot these shifts early will be the ones that will be able to maintain (and improve) their presence in AI search. If you want to see which sources Claude cites for your category, you can start a free Writesonic trial and check your own prompts.
How We Ran The Study
For anyone who wants to check our work, here's the setup. We ran 50 prompts across 16 categories through both search modes on Opus 5.5. Then we re-ran them on Sonnet 5.5 and Haiku 5.5 without specifying a search mode.
We added a head-to-head against Brave's and Parallel's search APIs and captured screenshots of the raw search data behind each answer. Data was captured on October 16th, 2026.
The study has limits. It's one account and one capture date. We haven't tested Fable 5.1 or older 4.x models yet. Two follow-ups are next on our list.
- First, a multi-account sample to tighten the October 5th date.
- Second, a category-level cut of which sources Claude trusts for SaaS, finance, healthcare, and more.
Sign up for our newsletter (below) or follow us on LinkedIn to stay updated.
The #1 brand in ChatGPT changes 52% of the time.
Get the report and find out why.
Director of Growth Marketing
Cody Slingerland is the Director of Growth Marketing at Writesonic. He leads SEO and AI search (AEO) strategy. He has over 12 years of search experience, helping companies like Zapier, Instapage, Nutshell, CloudZero, and others improve their visibility and position in organic search.
Data & Infrastructure @ Writesonic
Tarsh builds the data pipelines behind Writesonic's AI visibility research, including the AI Ads Index. His work focuses on real-time measurement of how AI platforms cite, mention, and rank brands.
