Made for advertising is used as an insult and it is a classification with commercial consequences. A domain that lands on the wrong list stops receiving bids from a large share of demand, and getting off that list is considerably harder than getting on it.
Having spent an uncomfortable amount of time on the wrong side of this, here is what appears to drive it.
Ad density relative to content
The most consistently cited signal and the easiest to measure automatically.
The ratio of ad slots to actual content, the proportion of the viewport occupied by advertising on load, and the number of ad calls per page view.
What crosses the line is not a fixed number, and the pattern is clear enough. Pages where advertising occupies more space than the article, where a reader has to scroll past multiple units to reach the second paragraph, and where the unit count keeps rising as you scroll.
Infinite scroll that loads new ad slots without loading meaningfully new content is a specific pattern that gets flagged.
Traffic acquisition
The second major factor and the one publishers find hardest to hear.
Sites whose traffic is overwhelmingly paid, particularly from cheap content recommendation and social arbitrage, get scrutinised in a way that sites with direct and organic traffic do not.
The logic from a buyer's perspective is straightforward. If a publisher buys traffic for less than it sells impressions for, the business is arbitrage, and the content exists to carry the ads rather than the reverse.
This is why the traffic mix matters as a signal independent of anything about the content itself. A site with the same pages and a healthy organic share is treated differently.
Content quality signals
The part that is hardest to characterise and the part that has changed most recently.
Classifiers look at whether content is original, whether it is substantive, whether it appears mass-produced, and increasingly whether it appears machine-generated.
The patterns that get caught are recognisable. Very high publishing volume with uniform structure. Articles that are recombinations of widely available information with nothing added. Uniform length, uniform heading structure, uniform tone across hundreds of pieces.
Templated headline formats repeated at scale. Comparison tables that appear in every article regardless of subject. The same call-to-action blocks in the same positions.
None of these individually proves anything. Together, at scale, across a domain, they form a signature that is easy to detect.
User behaviour
The signal that is hardest to fake and that ties the others together.
Time on page, scroll depth, bounce, and whether anybody ever returns.
A site where the median visitor arrives from a paid link, spends eleven seconds, and never comes back is telling the buyer something regardless of what the pages contain.
This is also why the fix is not cosmetic. Reducing ad density on a site nobody wants to read changes the density metric and not the behaviour metric.
What happens once you are on a list
The commercially serious part.
Major buyers maintain exclusion lists, and industry bodies have published frameworks for identifying this inventory. Once a domain is excluded by a large buyer, that demand does not return automatically when the site improves.
What remains is remnant demand at low rates, which is the outcome most publishers in this position describe — traffic holding up, revenue per thousand impressions collapsing.
Getting removed generally requires demonstrating change over a period, and in some cases direct engagement with the parties maintaining the lists. It is slow.
What actually seems to work
From what I have seen work rather than from what gets recommended.
Reducing ad density genuinely, not by moving units below the fold but by having fewer of them.
Shifting the traffic mix, which means building something that produces direct and organic visits, which is slow and is the only durable answer.
Publishing less and better, which reduces the volume signal and improves the behaviour signal at the same time.
And removing or improving the worst existing pages rather than only changing what is published going forward, since classifiers sample the domain rather than the newest content.
That last point is the one publishers most often miss. Adding good material to a large corpus of poor material changes the average slowly, and a sampled assessment is likely to hit the old pages.
The uncomfortable framing
The useful way to think about it, which took me a while to accept.
The classification is not really about content quality in an aesthetic sense. It is a proxy for whether the audience is worth reaching.
An advertiser paying for attention wants attention that exists. A page nobody reads properly delivers an impression and not attention, and the systems have got better at telling the difference.
Which means the question to ask about any page is whether somebody would be glad it exists, and whether they would come back. Every other signal is downstream of that.