Web Search with Highlights: 95.17% on SimpleQA, 35% Cheaper Than Claude's Built-In Search.

Chief Product Officer

The You.com Web Search with Highlights delivers our best SimpleQA score, 95.17%, a 16.5% relative lift, at no added cost (the same $5 CPM). A separate Braintrust evaluation also found that it cut the total cost of queries 35% versus Claude Sonnet 5's built-in search and 11% versus GPT-5.6 Terra's.

1. What Agents Really Need From a Search Result

Your agent doesn't click links. It reads whatever you hand it, so the payload should carry the answer itself.

Here is a real response to what was nvidia's q2 earnings?, showing the top result from NVIDIA's own press release.

*Same query, same index. Left: keyword-matched snippets. Right: passages selected for this specific question.*

Default, the snippets array:

"snippets": [
  "NVIDIA benefited from a $180 million release of previously reserved H20 inventory, from approximately $650 million in unrestricted H20 sales to a customer outside of China. For the quarter, GAAP and non-GAAP gross margins were 72.4% and 72.7%...",

  "Non-GAAP Measures To supplement NVIDIA's condensed consolidated financial statements presented in accordance with GAAP, the company uses non-GAAP measures of certain components of financial performance..."
]

With extraction_mode: "highlights", the contents.highlights array:

"contents": {
  "highlights": [
    "# NVIDIA Announces Financial Results for Second Quarter Fiscal 2026
- Revenue of $46.7 billion, up 6% from Q1 and up 56% from a year ago
- Data Center revenue of $41.1 billion, up 5% from Q1 and up 56% from a year ago",

    "NVIDIA (NASDAQ: NVDA) today reported revenue for the second quarter ended July 27, 2025, of $46.7 billion, up 6% from the previous quarter and up 56% from a year ago..."
  ]
}

Run it against your own key and diff the two responses:

curl -X POST https://ydc-index.io/v1/search \
  -H "X-API-Key: $YDC_API_KEY" -H "Content-Type: application/json" \
  -d '{"query":"what was nvidia'\''s q2 earnings?","count":5,"extraction":{"extraction_mode":"highlights"}}'

Drop extraction from the body and run it again. That's the whole difference.

2. How Highlights Works

*NVIDIA's earnings page, ranked and extracted.*

You.com Web Search with Highlights runs per request. Each query triggers a fresh extraction pass against the page, so passages are selected against what was actually asked. That's the NVIDIA page from Section 1: the revenue and Data Center figures survived because they answered the question, and the H20 inventory and non-GAAP boilerplate didn't, because they didn't.

Extraction works by selection. A model identifies which spans of text on the page already answer the query and returns those spans verbatim, with any numbers, dates, or figures exactly as they appeared in the source.

Tables return with their headers intact. Code blocks keep their original formatting and language. Sentences from the same section that both matter get merged into a single passage.

Candidate passages are ranked, and the highest-ranked ones are returned first. This means a short, relevant span still surfaces even in a document where it’s a small fraction of the total text.

3. Highlights Improves Accuracy

We ran three benchmarks against all three extraction modes.

	Snippets	Highlights	Full page
SimpleQA	81.68%	95.17%	88.98%
FRAMES	80.58%	82.04%	82.77%
RetrievalQA	63.91%	66.93%	63.77%

You.com Web Search with Highlights leads on SimpleQA and RetrievalQA, the two benchmarks built around single-fact lookups rather than multi-hop reasoning. On FRAMES, full-page extraction edges it by 0.73 points, inside the run-to-run variance we see on that benchmark.

That accuracy gain isn't free: cost per question rises about 11% versus Snippets, since Highlights sends more text to the answering model. But cost per correct answer, which divides cost by accuracy, still comes out about 5% lower — the extra spend buys more than it costs.

Against the field, that accuracy holds up, and it comes with a latency advantage the table above doesn't show. On SimpleQA, three systems clear 95%. For each competitor we're showing their best-performing configuration, not necessarily the one most comparable to Highlights, so the comparison favors them on setup.

System	Accuracy	p50 latency
You.com + Highlights	95.17%	695ms
Exa (full page)	95.47%	1337ms
Parallel (advanced)	95.33%	2098ms

Accuracy is a wash, within 0.30 points across all three. Latency isn't: You.com answers in roughly a third of the time it takes Parallel, and about half the time Exa takes.

4. Highlights Cuts Total Cost Across Every Tested Model, Up to 35%

Braintrust, as part of an independent evaluation, ran You.com Web Search with Highlights against the search tools bundled into the OpenAI and Anthropic APIs, on 1,329 questions, with cost including inference and search.

Model	Search	Accuracy	Cost per query	Latency
GPT-5.6 Terra	You.com + Highlights	84.82%	$0.0417	8.64s
GPT-5.6 Terra	Built-in	81.16%	$0.0467	12.04s
Claude Sonnet 5	You.com + Highlights	79.23%	$0.0747	11.48s
Claude Sonnet 5	Built-in	78.78%	$0.1148	12.74s

Claude Sonnet 5 costs $0.0747 per question with Highlights, against $0.1148 for its own built-in search — 35% less, at slightly higher accuracy (79.23% vs. 78.78%) and 1.26 seconds faster. At 100M queries — a volume we're using for illustration, not Braintrust's own figure — that difference is $4.01M

GPT-5.6 Terra costs $0.0417 per question against $0.0467 built-in, 11% less, and 3.66 points more accurate.

Cost per correct answer, which divides cost by accuracy, drops 35% for Claude and 15% for GPT, since Highlights is also getting more of the questions right on the same or lower spend.

Cost per correct answer tells you what a wrong answer costs. It doesn't tell you why one happened.

5. Unbundled Web Search Provides Better Observability Than OpenAI and Claude’s Built-in Search

Here's why. Built-in search hands back the queries it ran and the pages it cited. It does not hand back the passages the model reads, so a wrong answer arrives with no way to see which step produced it.

Braintrust's evaluation included a question asking how many service games Carlos Alcaraz lost across two tournament runs combined, requiring the model to add two separate numbers together.

The You.com configuration returned one passage establishing the first value at 24 and another establishing the second at 22, and the model added them to answer 46. The runs that missed never surfaced a passage supporting 24, used 22 for both values, and answered 44.

Every run did the arithmetic correctly, and you can only know that because the passages are in the trace. See the full example in Braintrust's writeup →

6. When Not to Use Highlights

Highlights adds about 160ms over Snippets on single-shot lookups. Under a hard latency budget on simple Q&A questions, Snippets remain the right default and they haven't changed.

For pages that move faster than an index refreshes, extraction_mode: "full_page" fetches live rather than serving from cache. FRAMES is the benchmark where that shows up: its questions are multi-hop, meaning the agent needs a wider slice of context to chain facts across several pages, and full page's larger payload has room for that where a handful of tightly-scoped passages doesn't. Full page beats Highlights there by 0.73 points.

7. Ship It

You've seen the payload, the accuracy, the cost, and the trace. The only thing left is to run it against a query that matters to you.

It's one field.

POST https://ydc-index.io/v1/search
{
  "query": "<query>",
  "count": 5,
  "extraction": { "extraction_mode": "highlights" }
}

Passages arrive in contents.highlights. Nothing else about your integration changes.

New accounts get $100 in credits, which covers every benchmark in this post end to end. The harness that produced these numbers is public, so point it at your own questions and see for yourself:

https://github.com/youdotcom-oss/web-search-api-evals

Read the docs