How do you research competitors with Claude? Ads, pages, positioning and AI answers

By Andrej Ruckij · October 29, 2026

Short answer: Claude won’t find you a hidden data source. It does the slow part after you’ve collected what’s public: describing a competitor’s ad precisely enough to rebuild its structure, reading pages that block scripts, sorting competitor messages into crowded and open, and checking who AI assistants cite in your category. Start with the public ad libraries (Meta’s needs no account), then work through the steps below, each with an output you can check. None of it shows a competitor’s results, and for ordinary commercial ads no ad library shows spend.

The usual prompt for this is “here’s a competitor’s website, give me a SWOT”. It produces a tidy table, and in my experience the table rarely changes a decision. What follows is put together from runs on my own products, a demo project and one client category. Each step has a real output and a limit.

Step 1: Collect the ads from the public libraries

Every big ad platform now has a public library of running ads. Two things matter more than which one you open first.

• Spend isn’t there. Meta shows spend ranges for political and issue ads. For ordinary commercial ads, none of the libraries (Meta, Google, TikTok, LinkedIn, Microsoft) shows what an advertiser spends.
• EU-targeted ads show the most. The EU’s Digital Services Act (Article 39) makes large platforms publish targeting and reach for ads aimed at the EU. TikTok’s library covers only ads shown in the EEA, Switzerland and the UK. LinkedIn shows targeting and impressions only for EU-targeted ads. If a US competitor also runs EU campaigns, the EU view is often the only public record of how it segments its audience.

On Meta, go to facebook.com/ads/library, pick a country, and search by page name or keyword. You see every ad a page is running now. Stopped ads stay visible only if they ran in the EU (about a year after the last view) or were political or issue ads (seven years). Steps are in questions/how-to-access-meta-ad-library, filters in questions/how-to-search-meta-ad-library, and history limits in questions/how-far-back-meta-ad-library.

Save what you find on the day you find it. Meta drops non-EU commercial ads once they stop, and the EU windows on Meta, TikTok and LinkedIn run about a year. After that, your own copies are the only history of a competitor’s campaigns. seo/meta-ad-library-api covers a weekly pull-and-compare, and competitor-analysis/ad-transparency-surfaces puts six platforms side by side.

The limit: a library tells you what ran, not what worked. An ad live for months hints that it pays for itself. Reach numbers are bucketed estimates, fine for comparing one competitor’s ads with each other and useless as measurement.

Step 2: Describe a winning ad precisely enough to rebuild it

Here Claude earns its keep. “Nice layout” is easy to say. Writing down which parts make an ad work, so you can rebuild the structure for your own product without copying it, is slow by hand.

For static ads I ask Claude for ten things: the layout grid, what the eye hits first and in what order, lighting, how much of each colour, typography, how the product is framed, setting, props, the feeling the ad promises, and the copy pattern. Then I ask for the same structure with the brand removed. Any Claude chat can do this with that list. We packaged it as a skill (ad-alchemy) so every ad gets the same treatment.

In April I ran it on a competitor’s meal-plan infographic, for a meal-planning app I run. It found a vertical two-column layout, five identical meal blocks, food photos cropped into dark-ringed circles, and curved arrows linking them. It named two things a remake has to keep or it stops reading like the original: the soft lighting inside the circles, and a headline mixing bold condensed type with one handwritten word. Then it wrote image prompts for five variants in our colours and language. The arrows were the least obvious find: without them the layout reads as five stacked food photos and loses the sense of a plan.

Video takes more structure. Our ad-replicator skill cuts a reel into shots, groups them into beats (one action each), and writes down what holds the beats together. One food reel I broke down was 21 seconds, 23 shots, 22 cuts. The camera never moved. Through most of the build, a hand came in from the right, added one ingredient and left, and the cuts fell while the hand was out of frame, which is why 20 of the 22 cuts were invisible. The more useful note: each addition changed colour or texture from the one before. A remake with higher-contrast ingredients could beat the original without changing a single shot.

This describes structure. Whether the structure is why the ad performs is something only your own test answers.

Step 3: Keep the structures as reusable templates

One reverse-engineered ad helps once. Stored as templates, several of them become a library you can fill with new messages and products. Ours are marked experimental until ad results come in. Our ad-template-foundry skill files each static ad as a layout with named slots, image rules and a note on when to use it, and for simple layouts it writes a small script that renders the template in 4:5, 1:1 and 9:16. The rule that keeps the library useful is merging: ads that share a layout become one template with different fillings. In the first run, ten source ads became four templates. More in marketing/static-ad-template-system.

Step 4: Read competitor pages that block bots

Ask Claude to read a competitor’s pricing page and you often get nothing. Many sites sit behind Cloudflare checks or render only in a browser, so a plain fetch returns a 403 or an empty shell.

Our page-snapshot skill opens the page in a real Chrome window, scrolls so lazy-loaded parts appear, saves the rendered HTML and converts it to text Claude can read. It takes one URL or a short list and follows no links. It’s built to stop and hand a captcha to you rather than try to solve it. To be straight about testing: the live page we’ve run it on blocked plain fetches with a 403 but showed no captcha, so that stop has been tested on sample pages, not yet on a real bot wall.

Use it on the pages that carry positioning: pricing, the homepage, “X vs us” comparisons, the features list. Save them with a date. Two copies of a pricing page three months apart say more than either alone. Pages behind a login you don’t have stay out of reach.

Step 5: Sort competitor messages into crowded and open

With the ads and pages in hand, the useful question is which messages are crowded and which nobody owns.

In June we mapped a DNA-skincare category for a client (competitor-analysis/dna-beauty-paid-social-whitespace). The map came from a sweep of competitor landing pages, teardowns and case studies; the ad libraries weren’t directly readable in that sweep, so its “open” calls carry a medium-confidence cap. “Personalised”, dermatologist authority and quiz funnels were crowded. Selling formulation plus a subscription was unmarketed in US paid social. The same sweep showed why: two premium players had already died in that lane, and the category fights a documented skepticism problem.

A crowded message can still be worth taking if you hold better proof. “Personalised” was everywhere, but a DNA test is stronger proof than a quiz, and the audience already distrusted quiz personalisation, so “quiz custom isn’t custom” stayed on the client’s list. Our target-audience-research skill ran that sweep and then graded every candidate angle against it; the grading method is in marketing/evidence-graded-audience-research.

The trap here is reading every empty lane as an opportunity. Look for the competitors who tried it and closed first.

Step 6: Check who AI answers cite

A competitor can win when a buyer asks ChatGPT, Copilot or Perplexity for options and still be invisible in Google rankings. Tracking this is called share of model (glossary/share-of-model): how often each brand is mentioned or cited when assistants answer your buyers’ questions.

A starting setup I’d suggest: ten buying questions, two or three assistants, the brands and sources that come back written down, repeated monthly. Answers change between runs and models, so read the direction over months, not one day’s ranking. Our ai-visibility-audit skill covers your own side: whether AI crawlers can reach and read your site at all.

Being cited and being visited are different things, and our own data shows how far apart. Microsoft Clarity’s AI citation report (a sample, covering Microsoft’s Copilot at least) counted 14,410 citations of our Meta Ad Library page over 30 days. About 30 sessions landed on that page, from all sources. Most queries it was cited for were misspellings of “meta ad library”: people looking for Meta’s tool. The lesson I’d carry over to competitors: read what the answers say about them as well as whether they’re cited.

What this method can’t do

• See spend or results. It works on what competitors show, not on what pays for them.
• See past the disclosure rules. Outside the EU every library shows less. X’s library is formally deficient: the EU’s €120M fine in December 2025 named it among other failings, and X is working through a fix-it plan the Commission accepted in July 2026.
• Replace talking to customers. Competitor ads show what competitors believe about your buyers. Reviews and win-loss interviews show what buyers believe; competitor-analysis/overview covers that layer.

If you’d rather have this done for your category, Primores reverse-engineers competitor creative as a service. The full method is here: https://primores.org/blog/b1p-ai-creative-reverse-engineering-complete-methodology/

Where to start this week

Pick your three closest competitors. Pull their running ads from the Meta library and save them with today’s date. Take the ad that has been live longest, give Claude the ten-point list from Step 2, and keep the one-paragraph structure it writes. Then ask two assistants your most common buyer question and note which brands they name.

Sources

• Meta Ad Library: the public library of running ads; no account required
• EU Digital Services Act, Article 39: advertising repositories for very large platforms
• European Commission press release IP/25/2934 (December 5, 2025): the €120M X fine, the first DSA non-compliance decision
• Primores runs: ad-alchemy (April 2026), ad-replicator (September 2026, demo reel), ad-template-foundry, page-snapshot, a DNA-skincare competitive sweep (June 2026, client anonymised), ai-visibility-audit. Internal run files.
• Microsoft Clarity, AI citations report for primores.org, 30 September – 29 October 2026 (our own data; sampled, Copilot at least)