How to Use AI Web Search With Citations

A plain AI answer is a guess shaped by training data. Switch on web search and the model goes out to live pages, pulls in what it finds, and shows you which pages it used. This guide covers how to use AI web search with citations in MultiChats: how to turn it on, how to read the source pills and the Sources list, how to pick a search provider and context size, and how to check the cited page yourself instead of trusting the summary. It is written for people who already chat with AI and now want grounded ai answers they can defend.

One thing up front: web search is a paid feature, and each search uses one message credit on a paid plan. On the web there is a small monthly free allowance so you can try it. On mobile it needs a paid plan to turn on at all. Plans are on the pricing page.

Step-by-step: turn on web search

The toggle lives in the composer, right where you type. Turning it on tells the model to search before it answers this message.

On web, click the globe button in the composer. When it is active the label reads Search. The globe appears on models with built-in web search (the Claude and Gemini models); to use search on any other model, first pick a default provider under Always use this provider for web search in Settings and the globe shows up on every model. On mobile, tap the + menu and choose Web search; the active pill also shows Search.
Type a question that benefits from fresh or sourced information. Search earns its keep on recent events, prices, specs, and anything you would otherwise paste a URL to check. For a static fact the model already knows cold, leave it off and save the quota.
Send the message. The answer streams in with a Sources section that shows progress: first Searched the web, then Fetched N websites as it pulls pages.
To turn it off again, click the globe (or the X next to it on web). Web search, image generation, and intent detection are mutually exclusive, so enabling one switches the others off.

Note that web search and the Create Image mode cannot run in the same message. Pick the one you need for that turn.

Reading the citations

This is the part most people skim, and it is the part that matters. A cited answer gives you two things to read: inline pills inside the text, and a Sources list at the bottom.

Inline citations show up as small pills with a favicon and the site domain, sitting next to the claim they back. Hover a pill (web) and you get a popover card with the page favicon, its domain, and the article title; click the card to open the actual page. When more than one source backs a claim, the pill shows a +N count you can expand. On mobile the same citations render, and tapping one offers an Open source link.

At the bottom of the message, the Sources list collects every page the model fetched, with lines like Searched the web and Fetched 3 websites. Read it as a map of where the answer came from. If a strong claim has no pill next to it, treat that claim as unsourced even when the rest of the answer is well cited.

Two quick habits make the pills useful instead of decorative. First, scan the domains before you read the prose. A claim backed by a primary source (a vendor's own docs, a court record, a research paper) is worth more than the same claim backed by a forum thread or an aggregator blog. Second, watch for a mismatch between the confidence of the sentence and the quality of its source. A flat, certain statement leaning on one shaky page is exactly where grounded ai answers go wrong, and the pill is what lets you catch it in seconds rather than after you have already repeated the claim.

Pick a search provider and context size

Open the sliders config next to the globe and you can choose which engine does the searching. The provider changes which results surface; with OpenAI Web Search selected you can also control how much page text it pulls back, which changes how deeply each page is read.

Provider	Good when you want
`OpenAI Web Search`	A general default that pairs tightly with the model's own answer style.
`Google Search`	Broad coverage and fresh indexing for mainstream, fast-moving topics.
`Anthropic Web Search`	Results wired to Claude when Claude is your chosen model.
`Perplexity Sonar`	A search-first engine if you prefer that style without a dedicated Perplexity app.

With OpenAI Web Search selected, a Context Size control appears next to the provider, set to Low, Medium, or High; the other providers manage this on their own. Low keeps the answer fast and reads less of each page; High pulls more page text in, which helps on dense or technical sources but costs latency. Medium is a sane starting point, and you can raise it when an answer feels shallow. If you came here looking for a Perplexity alternative, this is the closest control: same Sonar engine, plus 25+ other models in one place.

What a citation does and does not prove

A citation tells you a source was attached. It does not promise the page exists, that it says what the answer claims, or that the right passage was used. Researchers call the gap misgrounding: a model can state a fact correctly yet cite a page that does not actually support it. So a pill is a lead to follow rather than a verdict.

The risk is not theoretical. A Columbia Tow Center study (March 2025) tested eight AI search tools across 1,600 queries and found they answered more than 60% of source-identification queries incorrectly, with some engines citing fabricated or broken URLs in over half their responses. Counterintuitively, paid premium tiers scored worse on that test, because they answered confidently instead of declining. Grounding cuts down on AI hallucinations, but it does not remove them.

An earlier Tow Center test (November 2024) showed the same pattern from a different angle: given 200 traceable quotes, ChatGPT Search misattributed 153 of them, sometimes inventing citations or crediting a syndicated copy instead of the original. Search still beats a blind guess. The catch is that the attribution layer can be confidently wrong, so the pill earns trust only after you have opened it.

So treat the pill as step one. Open the source page behind it, read the actual passage, and confirm it backs the specific claim. For anything that matters, corroborate across two independent sources rather than one. The companion walkthrough on researching with AI and verifying sources goes deeper on lateral reading and tracing claims to primary documents, and our roundup of the best AI for research covers which models hold up under that kind of checking.

Web search checklist

Run down this list before you rely on a cited answer.

The globe shows Search (web) or the Search pill is active (mobile), so search actually ran.
The Sources list at the bottom names real pages, not just one vague entry.
Every load-bearing claim has a pill next to it; unsourced claims get flagged in your head.
You opened at least one source and confirmed the passage matches the claim.
For high-stakes facts (legal, medical, financial, exact numbers), you corroborated across two independent sources.
If the topic is contested, you re-ran it on a second model to see whether the answer holds.

Frequently asked questions

How do I turn on web search in an AI chat?

On web, click the globe button in the composer; it reads Search when active. The globe appears on models with built-in search (the Claude and Gemini models); set a default provider under Always use this provider for web search in Settings to get it on any model. On mobile, open the + menu and tap Web search. It applies to the message you send next, and you can turn it off the same way.

Does the AI show its sources?

Yes. Cited answers carry inline pills with a favicon and domain next to each claim, and a Sources list at the bottom of the message. Hover a pill on web to see the page title, then click the card to open the original page; on mobile, tap a pill and use its Open source link.

Is web search free?

It is a paid feature. On the web there is a small monthly free allowance so you can try it before subscribing. On mobile it needs a paid plan to turn on at all. Pro and Pro+ both include web search, and each search uses one message credit; see the pricing options.

Can I trust the cited sources?

Trust them after you check them, not before. A citation only means a source was attached; it does not guarantee the page supports the claim. Open the source, read the passage, and for anything important corroborate across two independent pages. When the stakes are high, compare answers across models as well, keeping in mind that models can share the same blind spots.

Wrap-up

Web search anchors an answer to real pages and hands you the receipts to check it. Turn on the globe, pick the provider that fits the job, read the pills and the Sources list, and open at least one source before you act on what you read. That habit turns a confident-sounding summary into something you can stand behind. Ready to search with citations, whichever model you chat with? See the plans and start grounding your answers.