Web Search in Chat

The AI Agent can search the internet inside a conversation. It uses web search when an answer depends on information that is not in your ad account, your client profile or the chat itself — a change in Meta’s advertising rules, a new Ads Manager feature, industry benchmarks for a niche or market, a brand or offer it has not seen before, or an outside cause of a performance swing (a holiday, a platform incident, a regulation). No confirmation is needed: a search reads public pages and changes nothing in your account.

When the Agent Searches on Its Own

• Meta policy and product news — “Is this claim still allowed?”, “What changed in Advantage+ this quarter?”, a deprecated placement or objective.
• Benchmarks and market context — typical CPM, CTR or CPA ranges for a vertical and country, seasonality, currency shifts.
• An unfamiliar brand, product or competitor — public information about the business behind an offer: positioning, price points, reviews, regulatory status.
• External causes of an anomaly — only after the account-side diagnosis found nothing internal.
• Verifying a claim you bring — “I read that Meta now requires…”.

The agent does not search for what it already has: campaign metrics, statuses and budgets come from your connected account, business facts from the client profile and team memory, competitors’ ads from the Meta Ad Library, and platform questions from this documentation.

How to Ask

Say it in plain language: “Find out whether Meta changed the rules for crypto ads in the EU”, “Check the typical CPA for meal delivery in Poland this year”, “Look up this brand before we plan the launch”. You can also add “verify this online” to any question. The agent phrases the queries itself; avoid including private account details or personal data in requests for public searches.

What You See

• While it searches, the activity line reads “Searching the web: ‘…’”, then “Found N sources”. Expanding the line lists the pages it found — title, domain and how old the page is.
• Statements taken from the web carry footnote markers in the text, and a numbered Sources list with titles and links appears under the reply. Markers and sources are kept in the conversation history and survive a page reload.
• The agent names the source type and date where it matters (“Meta’s Help Center article, updated in August”), separates facts it found from its own inference, and tells you plainly when a search returned nothing useful instead of guessing.

Cost

Web search is billed from the team balance as part of the turn, with the usual platform markup:

• a fee per search request (each query the model sends counts as one request, up to the platform limit per model call — 5 by default);
• the tokens of the results: the pages found are fed to the model as context (roughly 15–17k tokens per search) and are billed like any other input tokens.

The turn receipt shows a Web search line with the number of searches next to the model tokens, and the same number is recorded in the team’s AI spend journal. The balance check before a turn reserves enough for the searches one model call is allowed to run (5 by default), so a turn that cannot afford them is declined before it starts rather than cut short after the first search. Searches in the later steps of the same turn are paid from the team balance like any other tokens; the turn ends with the usual out-of-balance notice only when the balance itself runs out.

Switching It Off

Web search is a platform-level feature: an AdWitch administrator can disable it, change the maximum number of searches per model call (1–10) and block specific domains from the results. When it is off, the agent answers from your account data, its skills and this documentation, and says so when a question would need current outside information. A per-team switch is not available yet.