Web search for AI agents, with the source1 on every result.
Your agent asks a question. It gets back the exact sentence from the page, the link to that page, and the date the page was last read. Now it can quote the page and say where the words came from.
1 For developers. No public sign-up yet. We add people in small groups.
UserDoes the day pass cover the airport train?
Your agentNo. The day pass does not cover the airport train. Larkfield Transit says it “is not valid on the Airport Express (line A1), which needs a separate ticket.”[1]
- You sendA question in plain words
- You get backShort passages in JSON, each with its source
- You can limit toFresh pages, or a list of sites you trust
- We do notWrite the answer, or search private or paid pages
Every result carries three things.
An ordinary search result is a title, a link and a short teaser. That is made for a person who will click and read. An agent cannot click and skim. It needs the words themselves, and proof of where they came from.
The source page
The full address of the page, its title, and the site it sits on. Your agent can show the link to the person it is helping, so they can open the page and check.
We also tell you which heading the passage was found under. That helps a reader find the spot fast.
The exact passage
A passage is a short piece of the page, usually one to four sentences, that speaks to the question. We return it word for word, as it appears on the page. We do not reword it or shorten it in the middle.
Because the words are unchanged, your agent can put them inside quotation marks.
Two dates
Published is the date the page says it was written or last updated. Last fetched is the moment our system last opened the page and read it.
If a page does not state a date, we return null for published. We do not guess.
“Where did you read that?”
This is the question that breaks most assistants. Here is the same search, done two ways.
Results 1 to 2 of about 2. Sample.
Unlimited travel for one day. Buy in the app or at any station. Valid on bus and tram lines from first…
The Airport Express runs every 15 minutes from Central. Tickets are sold at…
The teaser stops before the part that matters. The agent must open each page, strip out menus and adverts, find the right lines, and work out how old the page is. Or it guesses.
Does the day pass cover the airport train?
No. Larkfield Transit says the day pass “is not valid on the Airport Express (line A1), which needs a separate ticket.”1 You can buy that ticket at the platform machines or in the app.2
- larkfieldtransit.example/fares/day-passpublished 2 Sep 2026 · checked 15 Nov 2026
- larkfieldtransit.example/airportpublished 12 May 2026 · checked 14 Nov 2026
The wording of the reply is written by your own model. We supply the quoted parts and the two footnotes.
Fig. 1. Left: what a normal search tool hands an agent. Right: what the agent can say when it has the passage, the link and the dates. All data is a sample.
One request in. Clean passages out.
You call SourcedSearch through an API. That is a web address your program sends a request to, and gets data back from. The data comes back as JSON, a plain-text format of names and values that every programming language can read.
# Ask a question. Only pages from the last 90 days, only two sites.
curl https://api.sourcedsearch.com/v1/search \
-H "Authorization: Bearer $SOURCEDSEARCH_KEY" \
-H "Content-Type: application/json" \
-d '{
"question": "Does the Larkfield day pass cover the airport line?",
"max_passages": 3,
"published_within_days": 90,
"sites": ["larkfieldtransit.example", "cityoflarkfield.example"]
}'
{
"question": "Does the Larkfield day pass cover the airport line?",
"filters": {
"published_within_days": 90,
"sites": ["larkfieldtransit.example", "cityoflarkfield.example"]
},
"passages": [
{
"rank": 1,
"text": "The day pass is valid on all city bus and tram lines from first service until 03:00 the next morning. It is not valid on the Airport Express (line A1), which needs a separate ticket.",
"source": {
"url": "https://larkfieldtransit.example/fares/day-pass",
"title": "Day pass · Fares · Larkfield Transit",
"site": "larkfieldtransit.example"
},
"position": {
"heading": "Where the day pass works",
"char_start": 1184,
"char_end": 1366
},
"published": "2026-09-02",
"last_fetched": "2026-11-15T06:12:00Z"
},
{
"rank": 2,
"text": "From 1 September the day pass also covers the river ferry between Quay Street and Northbank.",
"source": {
"url": "https://cityoflarkfield.example/news/ferry-day-pass",
"title": "Ferry joins the day pass · City of Larkfield",
"site": "cityoflarkfield.example"
},
"position": {
"heading": null,
"char_start": 96,
"char_end": 188
},
"published": "2026-08-21",
"last_fetched": "2026-11-14T22:40:00Z"
}
],
"left_out": {
"too_old": 1,
"no_date": 1,
"site_not_listed": 2
}
}
import os, requests
r = requests.post(
"https://api.sourcedsearch.com/v1/search",
headers={"Authorization": f"Bearer {os.environ['SOURCEDSEARCH_KEY']}"},
json={
"question": "Does the Larkfield day pass cover the airport line?",
"max_passages": 3,
"published_within_days": 90,
},
)
for p in r.json()["passages"]:
# Hand these three things to your model so it can quote and cite.
print(p["text"])
print(p["source"]["url"], p["published"], p["last_fetched"])
question- What you want to know, in plain words. No special search syntax.
max_passages- How many passages you want back. Fewer is cheaper for your model to read.
published_within_days- Skip pages older than this. Leave it out to search pages of any age.
sites- Only search these sites. Leave it out to search the open web.
char_start, char_end- Where the passage sits in the page text, counted in characters. Useful if you want to highlight it.
left_out- How many matching passages your filters removed, so an empty result is never a mystery.
The request, the response, the company and the sites are a sample. Field names may change before general release. People in early access are told before they do.
Between your question and our reply.
Fig. 2. The path of one request. The steps inside the yellow outline run on our side.
- 1
Your program sends a question
A normal sentence, the way a person would ask it. You can add a freshness limit, a list of sites, and how many passages you want.
-
Our side
- 2
We find pages that speak to it
We search our index of public web pages. If you set filters, pages that are too old or not on your list are dropped here, before anything else happens.
- 3
We pick the passages
Inside each page we look for the few sentences that deal with the question. Menus, cookie notices, adverts and footers are left behind. The passage is copied exactly.
- 4
We attach the source and the dates
Each passage gets its page address, title, the heading it was under, the published date if the page gives one, and the time we last fetched it.
- 2
- 5
Your agent writes the answer
You pass the passages to your own language model. It decides what to say. It can quote the text and list the sources underneath. That last step is yours, and we stay out of it.
Two switches: how fresh, and which sites.
Most wrong answers come from an old page or a page you would never trust. So these are the two limits you can set on any request. Try them on the sample below.
Limit the search
-
[1]larkfieldtransit.examplepublished 2 Sep 2026
It is not valid on the Airport Express (line A1), which needs a separate ticket.
-
[2]cityoflarkfield.examplepublished 21 Aug 2026
From 1 September the day pass also covers the river ferry between Quay Street and Northbank.
-
[3]larkfieldlocal.examplepublished 2 Oct 2026
Readers tell us the ticket inspectors on the A1 will not accept a day pass, whatever the app shows.
-
[4]tripnotes.examplepublished 9 Mar 2024
The day pass is great value and gets you everywhere, including out to the airport.
-
[5]larkfieldtransit.exampleno date on page
Airport Express tickets are sold at platform machines and in the app.
Look at card [4]. It is a 2024 travel blog and it is wrong today. With no limits it comes back, clearly dated, and your agent can weigh it. With either switch on, it never reaches your agent at all. A page with no date is treated as “not fresh” when the freshness switch is on. All cards are a sample.
Why two dates
and not one.
Fig. 3. A sample page over time. The striped gap is the time between our last read of the page and your question.
Published tells you how old the writing is
A price list from 2023 and a price list from last month can look the same. The published date lets your agent say “as of September 2026” in place of a flat claim.
We read this date from the page itself. If the page gives none, you get null and can decide what to do with it.
Last fetched tells you how old our copy is
Pages change after we read them. The last fetched time is the last moment we know the passage was on the page, word for word.
If that gap matters for your use, such as stock levels or live scores, your agent can open the link itself before it answers.
We bring the sources.
Some search tools hand back a finished paragraph. We chose to stop one step earlier, and here is the thinking.
What that means for the reader
The person using your product sees a quote and a link, and can check both in one click. Nothing sits between the web page and the quotation marks.
It also means SourcedSearch never states a fact in its own words. Every word we return was written by the page’s author.
Your model writes the answer.
What that means for you
You keep control of the voice, the length and the language of every reply. You can use any language model you like, and change it later, without changing your search.
When a reply is wrong, you can tell which part failed. Either the passage was a poor match, or the model misread a good passage. Those need different fixes.
For teams whose product has to show its sources.
A customer-help assistant1
Answer from your own public help pages and policy pages, and quote the line.
sitesA research agent2
Collect statements on a topic with a link and date beside each, ready for a reference list.
sourcepublishedA news or price watcher3
Ask the same question each morning and only read pages that are new since last time.
published_within_daysA coding assistant4
Pull the paragraph from the official documentation, not a forum post from years ago.
siteslast_fetchedAn assistant in a regulated field5
Keep a record of what was quoted, from where, and when it was last seen on the page.
textpositionlast_fetched
These are the uses we are designing for. We have not listed customers because we are in early access and will not name anyone without asking.
Three ways to give an agent the web.
These are three kinds of tool, described in general terms. Each is the right choice for some jobs.
| Classic search API | Answer engine | SourcedSearch[1] | |
|---|---|---|---|
| What comes back | Titles, links, short teasers | A finished paragraph, often with links | Word-for-word passages with link and dates |
| Who writes the answer | You, after opening the pages | The search tool’s model | Your own model |
| Can your agent quote it | Only after fetching and cleaning the page | Not safely, the text is reworded | Yes, the text is unchanged |
| Page age | Sometimes shown | Rarely shown | Published and last fetched, on every passage |
| Good when | A person will click through | You want a quick answer and trust the tool | The reader must be able to check the claim |
What SourcedSearch does not do.
Better to read this now than find out after a week of building.
It does
notwrite the answer.You get passages. You need your own language model, or your own code, to turn them into a reply.
It does
notsearch private pages.Nothing behind a login. No inboxes, shared drives, intranets or private documents.
It does
notget past paywalls.If a page shows only its first lines to the public, those lines are all we can return.
It does
notcheck whether a page is true.We tell you who said it, where and when. A page can still be wrong. The site list is your tool for that.
It is
notlive to the second.We return our last read of the page and tell you when that was. For data that moves by the minute, open the link.
It reads text, for now.
Words on web pages. Not the content of images, video, audio or scanned files. English pages come first in early access.
It does
notcover every page on the web.No index does. If we have nothing good, you get an empty list, not a weak match dressed up as a result.
Why build this now.
Web search was designed for people. A person sees ten links, picks one, reads, and judges the page with their own eyes. The judging happens in the reader’s head.
More and more, the reader is a program. An agent reads the results, writes a reply, and a person sees only the reply. The judging step has gone missing unless the reply shows its sources.
Language models are good at writing a smooth sentence. They are less good at knowing whether that sentence is true today. Giving the model the real words, the link and the date is the simplest fix we know. So that is the one thing we are building.
The technical words, in plain words.
You may be reading this for a team, not writing the code yourself. These are the seven terms this page uses.
- AI agent
- A program built on a language model that can take steps by itself, such as searching the web, reading a file or filling a form. An assistant is an agent you chat with.
- API
- A door between two programs. Your program sends a request to a web address. Ours sends data back. No person or web page is involved.
- Citation
- A note that says where a statement came from: the page, the link, and the date. The footnote under an answer.
- Fetch
- To open a web page by program and read what is on it. “Last fetched” is when we last did that for a page.
- Index
- Our stored copy of the text of pages we have fetched, arranged so the right passages can be found fast. Like the index at the back of a book, but for the web.
- JSON
- The format of the data we send back. It is plain text made of names and values, like
"published": "2026-09-02". - Passage
- A short piece of a web page, copied word for word. Long enough to make sense alone. Short enough to quote.
If the page being quoted is yours.
A tool that quotes websites owes something to the people who write them. These are the rules we set for ourselves.
We follow your robots.txt
robots.txt is a small file on a website that tells automated visitors which pages they may read. Our crawler, the program that fetches pages, names itself plainly and obeys that file.
If you want your site left out, write to hello@sourcedsearch.com and we will remove it.
We quote short, and always with the link
We return short passages, not whole articles. Every passage carries the address of your page. There is no way to ask our API for the words without the source.
The aim is that a reader ends up on your page to check the quote, not that your page is replaced.
# Your site. Your rules. User-agent: our-crawler Disallow: /drafts/ Disallow: /members/ # Or leave us out completely: User-agent: our-crawler Disallow: /
Questions developers ask first.
Not here? Write to hello@sourcedsearch.com.
Is the passage really word for word?
Yes. We copy the text as it appears on the page. We tidy spacing and remove layout code, and that is all. We do not reword, merge sentences from different places, or cut words from the middle.
What if a page has no published date?
Then published is null. We would sooner give you nothing than a guess. You still get last_fetched, which is always present.
Can I use it with any language model?
Yes. The reply is plain JSON. You pass the passages into whatever model or framework you already use. We have no model of our own in the loop that writes text.
How many sites can I put on my trusted list?
In early access the list is sent with each request, as in the sample. If you need a long, saved list, tell us in the form. It helps us decide what to build next.
What happens when nothing matches?
You get an empty passages list and a left_out count. That tells you whether there was nothing to find, or whether your own filters removed what was there.
Do you keep the questions my users ask?
We keep request logs for a short time to fix faults and stop abuse. We do not sell them and do not use them for advertising. The details are in the draft privacy policy.
Can it search my company’s internal documents?
No. SourcedSearch covers public web pages only. For internal files you need a different kind of tool.
Is there a price?
Not yet. Early access is free while we test. Prices will be set with early users and announced before anyone is charged.
Give your agent something it can quote.
SourcedSearch is new. We would sooner work closely with a few teams than open the doors and guess. There is no price list yet, and early access is free while we test.
What you get
- An API key and the reference documentation
- The search endpoint shown in the sample on this page
- A direct line to the people building it
- Notice before any field name or limit changes
What we ask
- Tell us what you are building, in a few lines
- Send us the questions where results were poor
- Expect rough edges and request limits while we grow
- Do not build anything life-critical on it yet
Request limits and paid plans will be set with early users before general release. We will not charge anyone without telling them first.