How to Use AI to Analyze Competitor Ads (Prompts and Workflow)

The fastest way to use AI to analyze competitor ads is to stop asking it to "analyze" a single screenshot and start feeding it a batch: pull 15 to 30 live creatives from a competitor or vertical, paste the headlines and a short description of each image into an LLM, and ask it to group them by hook mechanism, claim type, and target emotion. A single ad tells you almost nothing. A batch tells you what a competitor keeps repeating, which is the part worth copying the structure of, not the specific wording.

Why single-ad analysis fails#

Media buyers new to this workflow usually screenshot one competitor ad, drop it into a chat tool, and ask "why does this work." The model will confidently invent an explanation, because that's what LLMs do when given one data point and asked for a pattern. The fix is volume: patterns only become visible across a set of ads, and the model is much more reliable summarizing what's common across twenty examples than theorizing about one. This is the same principle behind analyzing winning native ad creatives: individual ads are noisy, batches reveal structure.

The prompt workflow that actually works#

Here's a sequence that holds up across health, finance, ecommerce and software verticals:

1. Collect. Pull a batch of live creatives for a competitor, brand, or vertical, ideally ones that have been running for a while (longevity is a rough proxy for "still working," covered in ad longevity as a winning signal). Note headline, image description, and vertical for each.
2. Classify. Prompt: "Here are 20 ad headlines from [vertical]. Group them by the core mechanism: curiosity gap, direct claim, social proof, fear/urgency, story format. For each group, note what percentage of the batch it represents." This turns a wall of headlines into a distribution you can act on.
3. Extract the claim structure. Prompt: "For each headline, extract the specific claim being made (not the wording, the underlying promise). Are the claims specific (a number, a body part, a named product) or vague?" Specific claims are the ones worth studying; vague ones are usually low performers padding out the batch.
4. Compare against your own creative. Prompt: "Here are my current headlines and the batch of competitor headlines. Which mechanisms am I not using at all?" This surfaces gaps faster than a manual side-by-side.
5. Generate variants from the gap, not from scratch. Once you know which mechanism you're missing, ask the model to draft headline variants using that specific mechanism, anchored to your actual product, not a generic template.

What to feed the model (and what not to)#

The quality of this workflow depends entirely on the input. Screenshots of the ad unit alone (image plus headline) are enough for classification; you don't need the landing page yet at this stage. Once you've shortlisted a few structures worth digging into, extend the same approach to the landing page and reverse-engineer the full funnel from creative to offer, since a headline that works can still fail if the landing page doesn't deliver on the specific claim it made.

Don't feed the model your entire raw export and ask for "insights" with no structure. Vague prompts produce vague, generic output regardless of how good the underlying model is. The more specific the question (mechanism, claim type, emotional register), the more useful the classification.

Where the source data comes from#

This workflow is only as good as the ad set you start with, which is the part most guides skip. Manually screenshotting competitor ads from a feed is slow and incomplete, since native ads rotate constantly and you only see what your own device happens to get served. A competitive ad intelligence workflow built around a watchlist and a searchable archive, rather than manual screenshotting, gives the LLM a much bigger and more representative batch to work from. Programmatic access matters here too: pulling data through an API rather than copy-pasting screenshots means you can automate the "collect" step entirely and feed a fresh batch into your analysis prompt on a schedule, which is exactly what the native ad data API is for. For teams already living inside an LLM workflow, there's also a direct route: the native ad library MCP lets you query the live creative corpus straight from Claude or ChatGPT instead of exporting anything first.

A worked example structure#

Say you're researching a competitor in the hearing-aid space. A batch pull might surface headlines like "Struggling to Hear Clearly? Discover a Device Transforming Lives" and several variants on that same curiosity-plus-claim structure. Classifying a batch like this typically surfaces two or three dominant mechanisms (a health curiosity gap, a device-focused direct claim, and occasionally a story format), and you can quickly see which one a competitor is leaning on hardest by volume, not by guessing from one ad. That distribution, not any single headline, is the actionable output.

Common mistakes with this workflow#

• Treating the model's output as fact rather than a hypothesis. LLMs are good at summarizing patterns you feed them and bad at knowing whether a pattern actually correlates with performance. Cross-check any pattern the model surfaces against real longevity or spend signals before betting a campaign on it.
• Analyzing too small a batch. Five ads isn't a pattern, it's a sample of noise. Twenty to thirty gets you somewhere real.
• Skipping the landing page. A great hook analysis means nothing if you don't also check what the funnel does with the click, which is where pre-lander formats usually decide whether the click converts.
• Ignoring geo and vertical differences. A claim structure that dominates in US health creative won't necessarily translate to a different geo or a software offer; re-run the classification per vertical rather than assuming one pattern generalizes.

Using AI to track changes over time, not just a snapshot#

A one-time analysis tells you what a competitor is doing this week. The more useful version of this workflow runs on a schedule: pull a fresh batch every few days, feed it the same classification prompt, and diff the output against last week's. A shift in mechanism distribution (say, a competitor moving from curiosity-gap headlines to direct-claim headlines) usually means they've found something that's converting better and are doubling down on it. Catching that shift early is worth more than any single deep analysis, and it's the difference between building a competitor watchlist that actually informs your creative calendar and one that's just a folder of old screenshots nobody revisits.

This is also where the "vague prompt" failure mode gets expensive at scale. If you're running this classification weekly across several competitors, small inconsistencies in how you phrase the prompt will make week-over-week comparisons unreliable. Lock the prompt template down once it's producing useful groupings, and change only the input batch, not the instructions, so your trend lines are actually comparable.

What this workflow won't tell you#

Be realistic about the limits. An LLM can classify mechanism and claim structure reliably; it cannot tell you a competitor's actual conversion rate, their true spend, or whether an ad is still active because it's genuinely profitable versus running on autopilot inside a set-and-forget campaign. Longevity is a proxy, not proof, and even a well-structured prompt is only summarizing what you fed it. Treat the output as a strong hypothesis generator that narrows down what to test next, not as a verdict on what will work for your offer.

Turning analysis into action#

The point of this whole exercise isn't a report, it's a shortlist of two or three claim structures you're not currently using, tested against your own product honestly. Once you have that shortlist, the AI's job shifts from analysis to generation: draft headline variants in the missing mechanism, keep the claim specific, and launch small. If you'd rather skip building the collection pipeline yourself, browsing a searchable, always-updating archive of live creatives inside OpenAdLibrary's ad intelligence platform gives you the batch to analyze without screenshotting a single ad by hand.