The copy behind winning Meta ads: a two-step pull in Claude

You ask the connected chat for your best ads and the copy they share. Back comes a confident list of hooks — urgency, social proof, a question in the first line. Then you open two of the “winners” in Ads Manager and the text is not what the summary described, because the table it reasoned from never contained any text.

That is not the assistant lying so much as filling a gap. The fix is to know which tool holds which half of the question, and to ask for the halves separately.

Two tools, one join key

ads_get_ad_entities is the tool for spend, results, click-through and every other metric. Asked on 2026-09-15 for body, title, headline, primary_text, message, description, call_to_action and object_story_spec, its field catalogue returned every one of them in unknown_fields. What it does return at ad level is creative_id, described as “the unique identifier of the creative used in the ad” — and marked neither filterable nor sortable.

The copy sits in ads_get_creatives. Its schema lists the readable fields: body, title, link_url, call_to_action_type, child_attachments and a handful of media and identity fields. It also explains the naming: body is what Ads Manager calls primary text, title is the headline.

The trap is in the listing. Called without IDs, the tool states it returns only id, name, account_id and status, and warns the assistant, in capitals, against reading those gaps as empty fields. Copy only arrives when you pass the creative IDs or name the fields. The per-ad alternative is ads_get_ad_preview, which returns body text, headline and CTA button alongside the rendered image — one ad at a time.

Ranking comes first and it has its own trap: several of your top ads can share one creative, so group by creative ID before reading any copy, or one piece of text gets counted as a pattern.

Splitting a question across the tools that hold its halves is a habit the Claude Ads Operator course drills, once the connector is attached properly.

Asking for it so the join actually happens

Step one: at ad level, pull amount spent, results, cost per result
and creative_id for last_30d. Group rows that share a creative_id and
sum them. Show me the table and stop.

Step two: for the creative IDs I pick from that table, fetch body,
title and call_to_action_type from the creative tool, by ID. Quote the
text exactly as stored. Do not summarise it or name patterns yet.

The pause between the steps is the point. You choose the winners from real rows — and ideally not by cost per result on tiny spend, which is how the wrong ads get crowned. Only then does text enter the conversation, verbatim, where you can read it before anyone tells you what it has in common.

A vendor guide shows the shape people reach for. Supermetrics’ blog (Andrea Linehan, 28 May 2026) suggests asking Claude to “break the data down by individual ad” and, for each ad, “also pull the headline and primary text”. That prompt is written for Supermetrics’ own connector, and this post has not tested it there. On Meta’s official connector the same sentence is two tools, and “cost per acquisition” is not a field name the catalogue knows — cost_per_result is, which is why a metric can come back as not existing.

Where this breaks

Multiple text options. An ad built with several primary texts or headlines stores them in what Meta’s AdCreative reference calls asset_feed_spec: “an asset feed with multiple images, text and other assets used to generate variations of an ad.” That field is not on the creative reader’s list, and ads_create_creative takes one message and one headline, not a set. What body returns for such an ad is not documented on the tool.

Meta does publish per-text reporting for these ads. Its Insights breakdowns page lists body_asset, title_asset, description_asset and call_to_action_asset among the dynamic creative asset breakdowns, and says they “only support a limited set of metrics”: impressions, clicks, spend, reach, actions and action values. None of those names resolved in the connector’s catalogue on 2026-09-15. Whether the entity tool accepts them as a breakdown anyway was not tested, so treat it as unknown rather than unsupported.

Text Meta changed. ads_create_creative describes its Advantage+ Creative switch as automatic per-impression optimisations including “text improvements”. Where that is on, the stored body is what you wrote, not necessarily what every person saw. A pattern drawn from stored text is a pattern in your drafts.

Video. The AdCreative reference describes body as “not supported for video post creatives”, so an empty body on a video ad can mean the text lives elsewhere, not that the ad had none.

The metrics table never contained the words. Anything said about them before the second call is a guess.

• Asking for patterns in one breath. The assistant will produce hooks and angles whether or not text was ever fetched. Make it quote the stored copy first.
• Counting a shared creative as several winners. Five ads on one creative are one piece of copy, not five votes for it.
• Crediting copy for a creative win. The same text on a different image is a different ad. Look at the preview before deciding the words did the work.
• Reading a multi-text ad’s body as its copy. Until you have checked Ads Manager, you do not know which option ran.

Try this tonight Run step one for your top ads by spend, then fetch the stored text for the three biggest creative IDs. Open the same three in Ads Manager. If the text matches and none uses multiple text options, the two-step pull is safe for your account and you can keep using it. If any differ, note which ones: those ads need judging from the preview, not from stored copy.