Search Providers Don't See the Same Web

Changing search providers can uncover pages your usual provider missed. We sent the same search requests to 12 providers and compared the first five results. Some providers returned nearly matching lists; others had little in common.

The returned pages are like reading lists from two librarians. If both lists recommend the same five articles, the second list adds no new places to look. A different list gives you more articles to check, though a new article may still miss the answer.

Result overlap by provider

brave and you.com shared about 81% of their first five results, roughly four out of five pages. serper, firecrawl and tinyfish shared about 70 to 73%. Across different provider families, average overlap was about 10%.

Those percentages describe the pages returned for the searches we tested. The reading lists show what each provider recommended, not every page each provider could find.

What a second provider adds

On the 363 test questions, tavily found 244 known answers and exa found 237. Checking both providers' first five results found 279 answers: 35 more than tavily alone.

Across combinations from different families, a second provider added 5.3 percentage points on average to the rate of finding the known answer. Combinations from the same family added 2.8 points. Both gains are measured against the better provider alone.

Larger gains from stronger providers

For this comparison, a stronger provider found the known answer in its first five results for more than half the test questions. We split Figure 2's 55 different-family combinations by that threshold.

Combining two stronger providers added 6.6 percentage points on average over the better provider alone (29 combinations). The other 26 combinations, with at least one provider at or below 50%, averaged 3.8 percentage points.

What's next

Different result lists can help explain why a second provider finds answers the first missed. But low overlap alone is no guarantee of useful results: the larger average gains came from combining providers that also performed well on their own.

For us at Telem, the next step is to follow those extra pages into an agent's answer. Finding the right page is only part of the job; the agent still has to read the results and answer correctly.

Our next post looks at what happens when an agent asks another provider after a miss. We'll examine whether the extra search helps the agent reach the answer, and how much more the agent has to read.

Methods and setup

Search setup

We reused the search requests from our first post, sent to 12 providers using Telem's recorded default settings on 11 and 20 October 2026. The answer comparison used 363 questions with complete evaluation records. The overlap calculation used all searches with returned pages from both providers being compared, giving 313 to 364 searches per combination.

Provider families

We grouped providers mainly by similar returned pages. The Google-results family contains serper, firecrawl, tinyfish, tavily and serpapi; brave and you.com form another family. exa, perplexity, parallel, ceramic and seltz each form a separate family.

serpapi belongs to the Google-results family because SerpApi supplies Google search results, despite sharing only about 12 to 16% of its first five results with the other members. These are comparison groups, not evidence of shared infrastructure.

Overlap

For each search, we count shared page addresses among two providers' first five results. Result order does not matter, and common address-format differences are removed before matching.

Two shared pages give 40% overlap. The denominator stays five when fewer results are returned. Figure 1 averages the per-search percentages over searches with returned pages from both providers.

Hit rate

A known answer is the answer page or value recorded before checking results. An automated rule counts a hit when the known answer appears in a provider's first five results.

For tavily, $244÷363\times 100\approx 67.2\%$. This measures the presence of the known answer, not an assistant's final response.

Added answers

For two providers, the combined hit rate counts a search as a hit if either provider finds the known answer. Each provider keeps its own first five results.

exa and tavily together found 279 answers, compared with 244 from tavily, the better provider alone. The gain was $(279-244)÷363\times 100\approx 9.6$ percentage points.

Limits and supporting data

The findings describe the recorded questions and settings, tested on different days. For seltz, the 11 October audit recorded news as the default page collection; documentation checked on 23 October instead names web as the default. The test results describe the earlier setup.