Within a single week in October, three of the largest ad platforms announced or expanded tools that read an advertiser's website and turn it into ads.
- Meta, at Advertising Week New York on October 6, said it is scaling smart product recommendations this fall. According to Relevant Audience's breakdown of the announcements, they analyze an ad's image, text and landing page to identify the featured product, then add related products to the ad.
- X launched X Lift the same day. Social Media Today reports that it converts a business URL into a set of ad options, images and copy included, and can build the whole campaign, targeting and all.
- Google added previews of AI-written headlines and descriptions generated from the final URL, and on October 12 began lifting promotions from some advertisers' pages into their ads.
Separately, Meta is testing a feature that builds its own landing page from your site. Its in-product description, captured by Jon Loomer, warns that "people who tap your ad may see this instead of your website." Meta has not documented the test, so treat the details as provisional.
The website used to be where the ad pointed. It is becoming what the ad is made from. In March we asked who answers for an ad when Meta writes it from a URL. Seven months on, the same question applies to three platforms at once.
How true sentences become a false ad
The governance problem this creates is easy to miss, because "generated from your page" sounds a lot like "consistent with your page." It isn't.
A model that reads a page has to choose what matters. Meta's AI landing page, per Loomer's capture, highlights your key services and calls to action: a summary of the site, not the offer in a particular ad. X says Lift was trained on X ads, so it optimizes for what performs on X, not for fidelity to your page. Smart product recommendations add products to an ad based on a reading of the page, and the announcements don't say whether those products appear on the page the ad links to.
The familiar risks return in a new order:
1. Selection. The model quotes the line it found most persuasive, which is often the one with the most conditions attached.
2. Compression. Nuance on the page becomes a confident sentence in the ad.
3. Staleness. The page changed after the ad was generated. The ad didn't.
4. Substitution. In Meta's test, the page the user sees may not be your page at all.
None of this is malicious. It is what happens when copy is produced at machine speed from source material written for humans. Pages were written to be read whole, with exclusions nearby and context above the fold. Ad generators read them for fragments.
A hypothetical, but not an exotic one. An outdoor retailer's site says "free delivery on orders over €50" in the header, and "excludes bulky items" in the delivery FAQ. A URL-to-ad tool reads the homepage, finds the delivery line persuasive and writes "Free delivery on all tents." Tents are bulky items. Every component of that ad came from the retailer's own site, and the ad is still wrong. Nobody wrote the false sentence. It was assembled from true ones.
This is why "it came from our website" will not hold up as a defense for long, with customers or with regulators. It explains where the claim came from. It doesn't make the claim accurate.
Four shifts we expect from URL-to-ad tools
- URL-in, ads-out onboarding will probably become the default for smaller advertisers. LinkedIn's Draft with AI already proposes ad copy from the advertised URL, and X has now joined. We'd expect most self-serve platforms to start new advertisers this way within a year.
- Previews will likely be the platforms' answer to governance. Google's previews show up to ten sample assets before launch. That is useful, and it is a sample. The combinations that serve at query time, paired with whichever page the system picks, are not the ones you previewed. We cautiously predict more preview and approval screens, and very little continuous checking from the platforms themselves.
- Page hygiene will become performance work. If any sentence on a page can become an ad, then out-of-date offers, buried conditions and generic hero copy stop being content debt and start being ad copy. Sensible teams will review high-traffic pages with the rigor they apply to creative.
- Visual drift will grow alongside textual drift. Meta launched catalog video the same week, and video generated from static and catalog images adds motion, scenes and end cards the page never shows. Expect more ads that look like a different brand from the page they land on.
The mistake to avoid is treating "generated from the page" as proof of alignment with it. If anything it runs the other way: the more automatically an ad is assembled, the less likely anyone at the advertiser has read it next to its page.
Check the page the machine read
The question worth asking about this shift is short and uncomfortable: which version of your page did the machine read, and is it the version your customer will land on? Most teams can't answer either half.
The practical answer has three parts. First, accept that landing pages are now inputs, and govern them like creative: named owners, review dates, offer expiry. Second, ask the platforms for the records they already hold: which generated assets served, against which page, and when. Some of that exists in asset reports today; much of it doesn't yet. Third, check outputs against inputs continuously. AdAlign puts each ad beside the page it links to and gives the pair an alignment score out of 10 across four dimensions, Visual Match, Message Match, Above the Fold Continuity and Tone Alignment, with every mismatch it finds written out as a sentence. When a model wrote the ad from the page, that side-by-side is often the only place a person sees both.
Run it at the moments drift happens: when a page changes, when the platform regenerates creative, and on a schedule in between, so alignment drift is caught in days rather than at the quarterly review. A good landing page experience used to mean a page that served the ad. Now it also has to be a page that is safe to quote.
The same pattern runs through this week's other pieces. Google now turns the coupons on your site into ads you never approved, and its localization beta translates your ad with a review step and your page without one. On TikTok, the page gives way to an in-app checkout and a shopping agent that wasn't in your ad approval meeting. And regulators judge the ad and its page together, which is the question in if the ad matches the page but not the checkout, was the ad honest?
Key takeaways
- Meta, X and Google now read advertisers' websites to generate ads, recommend products, extract offers or, in Meta's test, replace the page.
- Generated from your page is not the same as aligned with your page: models select, compress and go stale.
- Platform previews are samples, not monitoring.
- Govern landing pages as ad inputs, and score the resulting ads against them continuously.
Match the checking to your volume
If your creative volume is outpacing your ability to check it against your pages, pick the level that fits:
- Finding out whether this affects you: run a free audit on one ad. No call, no card.
- One team shipping changes weekly: AdAlign monitors the pages your live ads point to and alerts you when one changes. Every plan starts with a 30-day paid pilot on your own data.
- Several clients or brands: see how agencies run alignment checks across accounts.
Frequently asked questions
What are Meta smart product recommendations? Meta said at Advertising Week New York on October 6, 2026 that it is scaling smart product recommendations this fall. According to Relevant Audience, they analyze an ad's image, text and landing page to identify the featured product and then add related products to the ad.
What is X Lift? X Lift is an automated ad creation option X launched on October 6, 2026. An advertiser enters a business URL and X's AI generates ad options, including images and copy, and can build the full campaign including targeting. X says the system was trained on X ads.
What is Meta's AI landing page experience test? It is a test feature in Meta's advertising settings that uses AI to generate a visual landing page from an advertiser's website. Meta's in-product text says people who tap the ad may see this page instead of the website. Jon Loomer reported it in September 2026, and Meta has not published documentation.
How do I check AI-generated ads against my landing page? Compare each generated ad with the page it links to, not with the brief. AdAlign scores the ad and page pair on Visual Match, Message Match, Above the Fold Continuity and Tone Alignment, gives it one score out of 10 and lists the specific mismatches. Re-run the check when the page changes or the platform regenerates creative.