Diffbot Knowledge Graph for company news
Diffbot stores news articles as entities in its Knowledge Graph, with tags linked to the organizations they mention. That makes a company a node rather than a search string. It is priced in credits, with a free 10,000 a month, and it has not been run on the company-news benchmark.
Most of the APIs on the board take a question and return pages. Diffbot takes a graph query and returns records. This page covers what that means for company news, what it costs, and where the measured search APIs set the bar.
Benchmark last measured . 300 questions.
Articles as entities, organizations as nodes
Diffbot's Article type is described in its ontology as news, blog posts and article content known to the Knowledge Graph. Each article has tags, and the article extractor's documentation says the tags are produced by the Natural Language API and linked to the Knowledge Graph: a tag has a uri such as https://diffbot.com/entity/... and rdfTypes such as Organisation. The Organization type is the company record, with fields for location, employee count and so on. That is the link: article tag to organization entity.
Queries are written in DQL and sent to GET kg.diffbot.com/kg/v3/dql. The documented example filters organizations: type:Organization locations.city.name:"San Francisco" nbEmployees>5000. Article queries use the same language against the Article type: type:Article with field filters such as author, date and sortBy:date, as in Diffbot's own example type:Article author:"Marques Brownlee" sortBy:date. Each Article record carries a tags array, and each tag has a label, a uri to the linked entity, types, a mention count, a score and a sentiment. Two backends serve article queries: a fast index of the last six months by crawl date, and the full archive, which DQL routes to when the date filter reaches further back or when searchArchive=1 is set.
Two consequences. A company with a common name is not a problem, because the tag points at one entity. And the result is a record with fields, not a URL with a snippet, which is better for a pipeline and worse for a reader who wants to open the story.
Credits and what each call spends
- Free. $0, 10,000 credits a month, one thread. 429 once spent.
- Startup. $299 a month, 250,000 credits, overage $0.001 a credit, 25 threads.
- Plus. $899 a month, 1,000,000 credits, overage $0.0009, 100 threads, Crawl with 25 active jobs, three seats.
- Enterprise. Custom allotment, 100 or more threads.
- Per call. Extract one page: 1 credit, 2 through a datacenter proxy. Natural Language, one document up to 10,000 characters: 1. Knowledge Graph, export one entity record: 25; one facet query record: 100; enhance one record: 25, or 100 with refresh. Web Search, one search: 1. Crawl: 0.
The 25-credit export is the number to plan around. Ten article records for one company query is 250 credits; a thousand such queries is 250,000 credits, which is the whole Startup allowance. Diffbot's 1-credit Web Search is the cheap path and returns search results rather than graph records.
Vendor prices, limits and parameter names were read from the linked pages on 18 November 2026. They are the vendor's list terms, not measurements.
What the measured APIs do with the same question
Top five measured web search configurations on company-news questions
| Rank | Provider | Accuracy | Answer recall @5 | Mean search time | $ / 1k queries | $ / 1k correct |
|---|---|---|---|---|---|---|
| 1 | Exa (type=fast)POST /search · type=fast | 99.3% | 99.3% | 652ms | $7.00 | $7.05 |
| 2 | ContextPOST api.context.dev/v1/web/search · numResults=10 · highlights maxCharacters=1200 | 98.0% | 98.3% | 15.93s | $4.40 | $4.49 |
| 3 | Exa (type=instant)POST /search · type=instant | 97.7% | 97.3% | 398ms | $7.00 | $7.17 |
| 4 | PerplexityPOST /search · search_context_size=low | 97.3% | 98.3% | 1.38s | $5.00 | $5.14 |
| 5 | Linkup (depth=fast)POST /v1/search · depth=fast · outputType=searchResults | 96.7% | 94.7% | 1.57s | $5.00 | $5.17 |
Exa (type=fast) answers 99.3% of the benchmark's questions from the question as text, with no graph and no identifier.That is the lookup bar. Diffbot's argument is not that it finds the fact faster; it is that you already hold the company as a node, with its people, subsidiaries and funding events, and the articles should attach to that node. The benchmark does not measure that, and this page does not claim it either way.
Questions answered on this page
How does Diffbot find news about a company?
Through its Knowledge Graph. Diffbot extracts articles from the web, runs its Natural Language API over them, and links the tags it finds to graph entities, so an article about a company carries a tag whose uri points at that company's entity record, with a label, types, a mention count, a score and a sentiment. DQL, the graph query language, filters on the fields of the type you query; the endpoint is kg.diffbot.com/kg/v3/dql. type:Article queries run against a fast index of the last six months by default, or the full archive with searchArchive=1.
What does Diffbot cost?
Free: 10,000 credits a month on one thread, then 429 responses. Startup: $299 a month for 250,000 credits, overage $0.001 a credit, 25 threads. Plus: $899 a month for 1,000,000 credits, overage $0.0009, 100 threads, access to Crawl. Enterprise is custom. Exporting one Knowledge Graph entity record is 25 credits, extracting one page is 1, processing one document with Natural Language is 1, and one web search is 1.
So what does one company-news query cost in credits?
It depends on the backend. A Knowledge Graph export is 25 credits per entity record returned, so ten articles is 250 credits, or about $0.25 at the Startup overage rate. A web search is 1 credit. The two return different things: graph records with linked entities, or search results.
Has Diffbot been benchmarked?
No. The free plan's 10,000 credits cover the 300 questions only through the one-credit search backend, not the 25-credit graph export, so a fair comparison would have to pick a backend first. It has not been added to the board.
Diffbot or a search API for company news?
For one factual question, the measured web search APIs: Exa (type=fast) answered 99.3% from plain text at a per-query price a long way below 25 credits. Diffbot's case is the graph: when the company is a node you already hold, with its subsidiaries, people and funding events, and you want the articles attached to that node rather than the pages that mention a string.