AI Ad Creative Analysis: Hook, Body and Close on Revenue

Below: why CTR misleads, what the API returns, transcription, the tagging template, the join to revenue, a worked example, and how to run it with Claude.

Click-through rate measures one thing: how many people who saw the ad were curious enough to tap it. Curiosity is cheap to produce. A free offer, a shocking number, a vague promise to "10x your leads" all raise it. None of them says anything about whether the person who tapped has the problem you solve and the budget to pay for it.

• CTR rewards attentionThe hook does almost all the work. A broad hook reaches more people and more of them click, qualified or not.
• Cost per lead rewards easy formsAn ad that sets low expectations converts more visitors into leads. Whether sales can use them is invisible to the platform.
• Revenue rewards the right promiseThe ad that names the buyer's real problem and asks for a real commitment often has a higher CPL and the lowest cost per customer.

The platform's bidding makes this worse: optimise on leads and it finds more people who click and fill forms, which is the same signal the CTR ranking already over-weights. The fix on the delivery side is to send CRM stages back to Meta and Google. The fix on the analysis side is this post.

Before anything can be tagged, you need the words. Meta's Marketing API describes each ad's creative with these fields (as of October 2026).

For video: the file and the captions, not the words

The Graph API's Video node carries source, which Meta describes as "a URL to the raw, playable video file", plus picture, description, name and a captions edge. There is no transcript field. Whether you can read source depends on the access token you use. So the spoken words of a video ad come either from captions you uploaded or from transcribing the file yourself.

Meta's ads MCP server has tools for this layer too: ads_get_creatives, ads_get_ad_videos, ads_get_ad_images and ads_get_ad_preview, with ads_get_ad_entities for spend, impressions, CTR and conversions. Whether its video tool returns the source URL is not documented, so check what yours returns before you build on it. On Microsoft Advertising, the read only MCP server's GetAssets returns text, image and video assets with performance per association, and GetAssetGroups returns Performance Max headlines and descriptions. More in Meta Ads MCP and the ads MCP servers comparison.

Any speech-to-text tool will do, from a desktop app to a transcription API. What matters is what you keep from it.

1. Keep timestamps. The hook is defined by time (the first few seconds), so a transcript without times cannot be split reliably.
2. Add the on-screen text. Many ads carry the hook as a caption over silent footage. Note it with its timestamp, from the captions file or by watching.
3. Store one transcript per video id, not per ad. The same video runs in many ads; transcribe once and reuse.
4. Write down what the first frame shows. A face talking, a product screen, a chart. It is part of the hook even when no word is said.

• Hook

  The first line of text or the first seconds of video. Who it calls out and what it promises. It decides who stops.

• Body

  The middle: the problem, the mechanism, the proof. It decides who keeps reading and what they believe when they click.

• Close

  The ask and the button. What the person is asked to do and what it costs them: a download, a trial, a call.

The split is a judgement call, and that is fine as long as it is the same judgement every time. Write a one paragraph definition of where the hook ends for your formats (for example: the first sentence of text ads, the first spoken sentence or first caption of video ads) and apply it to every ad.

Free text cannot be grouped. Tags can. Fix a short list of values per column before you start, and let a value be "other" rather than inventing a new one per ad.

This is where an AI model saves hours. Give Claude the transcript and text fields of each ad plus the template with its fixed values, and ask it to fill one row per ad and quote the sentence behind each tag. Check a sample by hand; if more than a few rows are wrong, tighten the definitions rather than the prompt. A Claude Skill (a folder of instructions Claude loads for a task) is a good place to keep the template so every run uses the same one.

The tags are half the dataset. The other half is, per ad: spend from the platform, and leads, qualified leads, won deals and revenue from the CRM. The join needs the ad on every lead, which means one of two things.

• The ad id in the URLA UTM parameter (often utm_content) filled with the ad id or a name that maps to it, carried into a hidden form field and stored on the lead. Breaks on redirects and when people come back later without the parameters.
• The click id joined to the dealfbclid, gclid and the others kept with the visitor and resolved to the ad, then followed through the CRM. Survives a return visit days later. See click ids.

Then decide the revenue rules once: closed-won value or paid orders, never pipeline; which attribution model credits the ad; and a cut-off that leaves time for deals to close. A creative that ran last week has not had a chance to produce revenue yet.

Ranked on CPL, the result claim hook wins by half. Ranked on cost per qualified lead and on ROAS, it comes last, and the product demo hook leads. That reversal is the whole reason to do this work.

In one Claude chat, the platform connector supplies creatives and spend, and the LeadJourney MCP server supplies the CRM side per ad. The transcripts you made come in as files. A session looks like this:

1. "List every ad that spent in the last 90 days with its creative text, video id and spend." The Meta connector answers.
2. "Here are the transcripts by video id. Fill the tagging template for each ad and quote the line behind each tag."
3. "For the same ads and dates, give me leads, qualified leads, won deals and revenue." The LeadJourney connector answers from the CRM.
4. "Group by hook type, then by ask. Show cost per qualified lead and ROAS, with the counts beside them."
5. "Which tags show up in our best ads by qualified leads but not in our best ads by CTR?" This is the question that pays for the exercise.

The output is a brief for the next round of creatives, not an automatic change. Any edit to a live ad goes through the platform connector or Ads Manager, with you approving it. More prompts in Claude for Meta Ads.

LeadJourney does not analyse creatives, tag hooks or transcribe video. It supplies the half of the dataset the platform cannot: which ad each lead came from and what that lead became. One script, live in about 21 minutes, tracks every click, form fill and call server-side and first-party at 95%+ accuracy, keeps the click ids and joins each lead to its CRM deal in HubSpot, Salesforce, Pipedrive, Close, Attio, GoHighLevel, ActiveCampaign, Odoo, Zoho CRM or Microsoft Dynamics 365.

The result is spend, leads, qualified leads, won deals and revenue per campaign, ad set and ad, synced with Meta, Google, LinkedIn, Microsoft, Reddit, X, Taboola and Outbrain, under the attribution model you choose. The same numbers are in the MCP server for Claude or ChatGPT, which reads them and builds reports but never edits an ad. The CRM stages also go back to the platforms as conversions, so delivery learns from the ads that make customers.

• Which ad copy attracts leads that close
• Automated budget rules on cost per qualified lead
• Claude for Meta Ads: connector, skills and 25 prompts
• Meta Ads MCP plus your CRM: which campaigns made revenue
• Ads MCP servers compared
• Why your ad platform's MCP gives Claude the wrong CPL
• How to increase the quality of your leads in performance marketing
• Lead quality per channel and campaign