First-party research · September 2026
Claude doesn’t search before recommending a payments API
Asked “What is the best payments API for a SaaS platform?”, Claude answered from memory in three out of three runs. It had a web search tool available and chose not to use it. Across the 11 questions in our payments measurement, Claude skipped search on 5, and each of those 5 went the same way on every run.
That matters for one reason. If an engine doesn’t search, no website is cited, and no change a company makes to its site today can appear in that answer. On those questions, Claude’s recommendation comes from what it learned in training.
What we measured
On 25 September 2026 we ran the Cited Score payments question set (version 2): 11 buyer questions, each asked 3 times, of 4 engines — 132 calls. Every engine was queried through its API with web search available. The engines differ in whether searching is a choice:
| Engine | How search works in this measurement |
|---|---|
| ChatGPT (OpenAI API) | forced on every call |
| Perplexity | search-native; every answer is retrieved |
| Gemini | grounded on every call |
| Claude (Anthropic API) | tool available; Claude decides |
So of the four, Claude is the only one whose decision not to search can be observed. The other three are configured, or built, to always retrieve.
The result
| Question | Intent | Claude searched |
|---|---|---|
| What is the best payments API for a SaaS platform? | recommendation | 0 of 3 |
| Which payment provider should a marketplace use to pay out sellers? | recommendation | 0 of 3 |
| Best payment processor for a company expanding into multiple countries | recommendation | 3 of 3 |
| Which payments platform has the best developer documentation? | recommendation | 3 of 3 |
| Recommended payment gateway for a subscription billing business | recommendation | 3 of 3 |
| Best payments infrastructure for a fintech handling cross-border transfers | recommendation | 3 of 3 |
| How do the major enterprise payment processors compare? | comparison | 3 of 3 |
| What are the alternatives to Stripe for online payments? | comparison | 0 of 3 |
| Which payments API has the lowest transaction fees for high volume? | comparison | 3 of 3 |
| How does a payments API work for an online platform? | category | 0 of 3 |
| What should I look for when choosing a payment processor? | category | 0 of 3 |
Claude never split a decision: each question went the same way all three times. When it did search, it ran one or two searches per call.
Why this differs from what we saw before
In our embedded insurance measurement, Claude’s decisions not to search clustered on category-definition questions — the “how does X work” kind, where a search adds little. Payments is different: Claude also skipped search on the lead recommendation question, on a payouts recommendation, and on “alternatives to Stripe”.
Our reading, which is interpretation rather than measurement: embedded insurance is a young category with thin coverage in training data. Payments is mature and heavily written about, and Claude appears confident enough about the settled questions to answer them from memory. It searched where the answer depends on things that change or vary: fees, cross-border coverage, subscription specifics, an enterprise comparison.
What it means for AI visibility
For a payments company there are two ways into a Claude answer. Retrieved: Claude searches, finds your page and cites it — what content, structure and technical SEO can influence, and only on questions where Claude searches. Recalled: Claude names you from training data, with no source. On the lead question, the payouts question and “alternatives to Stripe”, recall is the only way in.
These are separate measurements, and the Cited Score keeps them separate: cited means your domain appeared in the engine’s sources; mentioned means your name appeared in the answer. Where Claude answers from memory a citation is not possible, so those calls are recorded as failed measurements, not as zeros. A zero would claim the company was invisible; what actually happened is that nothing was retrieved at all.
The practical consequence: on Claude, for a mature category’s most valuable questions, visibility is decided before the question is asked. It is shaped by how widely and consistently a company has been written about across the sources that feed training, not by what its own site says this week.
A side effect: it shows up in the bill
Search fees are the largest single cost of measuring AI visibility. This scan cost $1.75 against $2.28 for our previous full scan, and most of the difference is the 15 Claude calls that ran no search and so incurred no search fee.
Limits
One scan: one measurement of one category on one date. The consistency across runs is strong evidence the behaviour is systematic for these questions, not evidence that it holds for every question or category. API, not the app: we query Claude through Anthropic’s API with the web search tool enabled; the Claude app makes the same kind of decision, but its configuration is not identical. And this scan did not keep Claude’s from-memory answers. From the next scan onward they are stored, so we can report what Claude recommends when it does not search. They will never enter a score, because a citation measurement cannot be taken from an answer with no sources.
Measured for your company
The same measurement runs against any company in a covered category — your score per engine, and the competitors named instead of you. Pricing and access.
How often each payments company was named across the 115 successful calls is on the payments leaderboard. How the engines’ sources differ from each other is in AI engines do not share a citation layer. How the Cited Score is calculated, and why cited and mentioned are never combined, is in the methodology.
Questions or replication requests: [email protected]