Facebook Ad Library search through Claude: what comes back

You ask Claude what a competitor is running. Back comes a tidy list — page names, ad IDs, a couple of dates — and you realise you have learned nothing about their offer, their angle or their price.

That is not the model being lazy. It is the whole of what the tool handed it. The decision worth making before you start is which job you are doing. Reconnaissance — who is advertising in this category, since when, at what volume — is a job this payload can do. Creative research needs the words and the pictures, and the words and the pictures are the two things it does not carry.

What ads_library_search returns

Five inputs: search_terms, page_ids, countries, ad_active_status and ad_type. At least one of the first three is required, and asking without them gets a flat refusal rather than a default:

At least one of search_terms, page_ids, or countries must be provided.

When it does answer, every ad in the list has the same shape. Here is a real row with the identifiers cut out:

{
  "estimated_total_count": ...,
  "ads": [
    {
      "id": ...,
      "page_id": ...,
      "page_name": "Mind & Mood Daily",
      "ad_creative_link_title": "Feel More Like Yourself Again",
      "ad_creation_time": ...,
      "ad_delivery_start_time": ...,
      "ad_snapshot_url": "https://www.facebook.com/ads/library/?id=...",
      "currency": "USD"
    }
  ]
}

Eight keys per ad, and no fields parameter with which to ask for more. Both times arrive as raw epoch integers, so any calendar date you read in the chat is the model's arithmetic rather than something Meta sent.

Set that against what Meta publishes through the Ad Library API. The ArchivedAd reference lists ad_creative_bodies — the actual text of the ad — alongside ad_creative_link_descriptions, ad_delivery_stop_time, publisher_platforms, languages, target_ages, target_gender and target_locations. None of those are reserved for political advertising. None of them reach the chat.

Keyword search is looser than it looks

On 27 August 2026 I asked for ten active ads in the United States matching sleep supplement. Every row came back from novel-promotion pages, two of them between the lot, each carrying an identical link title inviting the reader to start reading. Nothing about sleep. Nothing about supplements. The same call for magnesium glycinate, minutes later, behaved exactly as you would want: supplement advertisers, on topic, ranked sensibly.

One term worked and one returned an unrelated category, on the same day, through the same tool. Treat the keyword path as a way to discover page IDs, and check what came back before a summary gets written on top of it.

What the Library holds, and what arrives in the chat

The last row is the one worth building a routine around. Everything above it is a reason to open a tab.

Knowing what a tool will not tell you is what stops a summary becoming a decision. See how the weekly routine is built →

How to get something out of it anyway

1. Search by page, not by keyword START HERE

   Run one keyword search to harvest page_id values, check by eye that the pages returned are actually your competitors, then throw the keyword away and pass those IDs in page_ids from then on. A page search is exact; a keyword search is a guess about how Meta indexed somebody's copy. The trade: you have to already recognise who your competitors are — and the keyword path that would have told you is the one you just stopped trusting.

2. Count advertisers, not ads

   Ask for a page of results and tally how many rows each page_name took. In my supplement search, ten rows resolved to three advertisers, and in the sleep search to two. That concentration is the signal the thin payload is genuinely good at: it tells you who is running enough volume to dominate an index they do not control. The trade: ad count is a poor proxy for budget. A brand running dozens of near-identical ads may be spending heavily, or may be testing badly, and this payload cannot separate the two.

3. Open the snapshot URLs for the handful that matter

   Every row carries ad_snapshot_url. For the ads that survive the first two steps it is worth the click — the rendered ad, the copy, the format — and you can paste what you find back into the chat. The trade: it is manual, and it is exactly the step the whole exercise was meant to remove.

The Ad Library indexes ads. Advertisers think in creatives. That gap is your result page.

• Treating the keyword search as a topic search. It matches text in ad creative, which is not the same as matching a subject, and one of my two probes proved it on the day.
• Reading the result count as market size. The response calls it an estimate, and it counts ads, not advertisers and not spend.
• Asking Claude to preview a competitor's ad. The preview tool works on ads in an account you have access to. For someone else's ad, the snapshot URL is the route — and yes, the connector reads creatives fine for your own account.
• Letting the model find "themes" in link titles. A link title is often a discount line or an empty string. A theme built from those is a theme about button copy.

Try this tonight Run one keyword search for your category, active ads only, in your country. Ignore every field except page_name, and count how many rows each page took. The one or two pages that dominate are worth an hour — open their snapshot URLs and read the offers. The pages that appear once are noise, and you have just saved yourself researching them.

The same de-duplication problem shows up inside your own account, where you at least have a key to group on: de-duplicating by creative_id covers that. The Ad Library payload has no such key. If you want the rest of what the connector exposes, the full tools list has it, and if you have not connected yet, start with the setup.