Made-for-Advertising Sites: The Legal Grey Zone That Beats Every Fraud Filter

Run a made-for-advertising site through every fraud tool the industry sells and it will pass. The traffic is human, so the invalid-traffic vendor clears it. The ads are on screen, so the viewability vendor scores it well. The domain is real, the ads.txt is valid, the seller is authorised. Every box is ticked.

And the advertiser's money has gone to a page that exists for no reason other than to hold ads, read by someone who arrived from a paid social post about a celebrity and left in nine seconds.

That is the problem with MFA. It is not fraud in the sense that anyone is being technically deceived, which is exactly what makes it harder to deal with than a botnet. You cannot filter it with a rule. You have to judge it.

It is also worth being precise about the accusation, because "MFA" gets thrown at any site somebody finds ugly. Nothing described in this article is illegal. No standard is being violated. In most cases no contract is being breached. The operators are buying media and selling media, which is a business the entire industry is built on.

The complaint is narrower and harder to legislate: the inventory has no reason to exist except to absorb ad spend, and it is being bought by machines that were never asked to have an opinion about that.

The recipe

MFA operations follow a pattern consistent enough to write down as a procedure, which is itself the tell — legitimate publishing does not converge on a single format this tightly.

1. Find topics that generate cheap clicks. Trending, controversial, emotional, or nostalgic. Celebrity health rumours, "you won't believe what happened to", listicles about the 1990s, mild outrage about something local. The subject matter is chosen for cost-per-click on social platforms, not for whether anyone needs it written.
2. Build minimal content around them. Recycled, rewritten, increasingly AI-generated, frequently barely coherent past the second paragraph. The text exists to justify the page, not to be read.
3. Buy traffic to it. Paid social on Facebook, Instagram and TikTok; content-recommendation widgets; search arbitrage. Wherever the cost per click is low enough that the arithmetic in step five works.
4. Load the page heavily. Eight to fifteen ad units, often including slow video. Infinite scroll that triggers more as the reader moves. Auto-refresh on short intervals. Units placed so they obscure the content. Thin text stretched across many screens to manufacture scroll depth, because scroll depth means more refreshes.
5. Check the spread. If revenue per thousand visitors exceeds the cost of acquiring them, the loop is profitable.
6. Reinvest and scale. The margin funds more traffic, which funds more margin.

Step four is where the reader experience is sacrificed, and it is sacrificed deliberately rather than through incompetence. Every element that makes an MFA page unpleasant — the density, the refresh, the layout that fights you — is a revenue optimisation. The site is working exactly as designed.

The refresh mechanic, specifically

Auto-refresh deserves its own paragraph because it is the single largest multiplier in the model and it is not, in itself, illegitimate. Refreshing an ad slot after a genuine interval, for a reader who is still present and still engaged, is normal practice and there is published guidance on doing it responsibly.

MFA takes the same mechanism and removes every constraint from it. Refresh intervals compress to the shortest the demand source will tolerate. Engagement gating — only refreshing for a reader who is actually there — is absent, because gating reduces impressions. The slot regenerates whether or not anyone is looking.

The result is that a single visitor, in a single session lasting under a minute, can generate the impression volume of a dozen genuine pageviews. That is where the arbitrage margin actually comes from. It is not that MFA sells impressions at a high price; it is that it manufactures a great many of them per human.

Why no single test catches it

Here is the property that makes MFA structurally different from every other category in this Journal: it defeats the specific tools buyers rely on, and it does so honestly.

Invalid-traffic detection asks whether a human was present. On an MFA site a human genuinely was — they clicked a real ad on a real social platform and landed on a real page. IVT vendors pass it because there is nothing invalid about the traffic in the sense IVT measures.

Viewability measurement asks whether the ad was on screen long enough to be seen. MFA pages are engineered for exactly that. Dense layouts fill the viewport, refresh cycles guarantee fresh impressions in view, and content stretched over many screens keeps the reader scrolling past units. MFA frequently scores better on viewability than a well-designed editorial page with two ad slots and a lot of text.

That sentence is the whole problem. The measurement layer was built to answer "was this real" and "was this seen". MFA answers yes to both. The question it fails is "was this worth buying", and nobody built a vendor for that because it is a judgement rather than a measurement.

What actually defines MFA, then

No single characteristic is disqualifying. Plenty of legitimate publishers run a lot of ads. Plenty use auto-refresh, responsibly. Plenty buy some traffic. MFA is a combination, and the anchor of the combination is the traffic source: the site has little or no audience of its own and depends on paid acquisition for essentially all of its visitors.

That is the line that separates an aggressive publisher from an arbitrage operation. An aggressive publisher has readers and monetises them hard. An MFA operation has no readers — it has a media buying account.

Why buyers keep buying it anyway

There is a temptation to read the scale figures below as evidence that buyers are careless. Some are. But the more useful explanation is structural, and it is worth stating because it predicts how the problem behaves.

Programmatic buying is largely automated optimisation against stated goals. A buyer sets a target — cost per thousand, cost per click, viewable rate, completion rate — and the system finds inventory that delivers it cheapest. MFA inventory is cheap, delivers reliably, passes verification and scores well on viewability. Against those instructions it is not a mistake. It is the correct answer.

Which means the problem is not that the machines are failing at their job. It is that the job they were given does not include "and the page should be worth being on", because nobody has a reliable way to express that as a number.

The scale

The ANA's programmatic transparency work in 2023 found MFA sites accounted for approximately 21% of impressions and 15% of spend in the campaigns it examined. The same work found the average campaign ran across roughly 44,000 websites.

Sit with the second number for a moment, because it explains the first. Nobody inspects 44,000 websites. No buyer has ever reviewed that list. It is assembled by algorithms optimising for cost and delivery, and MFA inventory is cheap and delivers reliably — which is precisely what those algorithms are told to find.

HUMAN's 2025 research put MFA inventory at over one trillion bid requests a month. At that volume MFA is not an edge case in the long tail — it is a substantial share of what the open programmatic market actually consists of.

The gap between those two figures is itself informative. MFA takes a larger share of impressions than of spend because the inventory is cheap — which is exactly why the algorithms buy it. Low CPM, high delivery, passes verification. From the optimiser's point of view it is the ideal inventory. From the brand's point of view it is a page nobody wanted to be on.

What happened after the ANA study

The 2023 findings did produce a response. Inclusion lists became more common, buyers started asking harder questions about supply paths, and several DSPs and verification vendors shipped MFA-specific classifications and avoidance controls.

The category did not disappear. The reason is the one that runs through everything in this Journal: MFA is profitable, cheap to produce and easy to relaunch under a new domain. Detection improving raises the operator's cost from near-zero to slightly-above-zero. That filters out the laziest and leaves the rest, which is the same shape every other enforcement story in this field takes.

What genuinely changed is that "we did not know" stopped being an available answer for a buyer. The data is published. The controls exist. Continuing to fund the category is now a choice rather than an accident, which is progress of a modest kind.

How to actually spot it

Since no single test works, the practical approach is to look at the thing MFA cannot hide: where its audience comes from.

The traffic-source fingerprint

A real publication accumulates an audience. Some people search for it. Some type the URL. Some arrive from a newsletter, or a bookmark, or another article. The referral mix is diverse and it is stable, because audiences change slowly.

An MFA site's referral mix is dominated by paid channels — content-recommendation widgets, social ads, search arbitrage — with organic search and direct traffic close to nothing. And it moves fast, because it follows whatever is cheapest this week rather than whatever an audience habitually reads.

What pre-bid systems look at

Modern pre-bid evaluation, as described in DoubleVerify's research on the category, combines several weak signals into a usable judgement:

• Historical traffic-source composition — and specifically abnormal spikes in paid social, which is what scaling an arbitrage campaign looks like from outside.
• Ad density relative to text — the ratio of ad units to actual words on the page, which is a fair proxy for what the page is for.
• Repetitive generative-AI text structures — the characteristic sentence rhythms and paragraph shapes of bulk-generated content.
• Domain age — MFA domains are disposable and frequently young, because the model tolerates burning a domain and starting again.

None of those is conclusive alone. A new domain is not evidence of anything; every legitimate site was new once. Together, with the traffic-source pattern as the anchor, they resolve into something you can act on.

What the site itself tells you, if you read it

Beyond the traffic data there are content-level tells that survive most attempts to disguise them, because fixing them costs money the model does not want to spend:

• No named authors, or authors with no history. A byline that appears on four hundred articles across nine unrelated subjects in six weeks is a template field, not a person.
• No corrections, no updates, no editorial policy. Real publications get things wrong and say so. A site that has never corrected anything has either never made a mistake or has nobody checking.
• Topic incoherence. Celebrity health, cryptocurrency, home remedies and local politics under one masthead is a keyword strategy wearing a publication's clothes.
• Content that ends rather than concludes. Bulk-generated text tends to run out rather than arrive somewhere, often mid-argument, often with a heading that has nothing under it.
• A contact page that contacts nobody. A form with no address, no company registration and no named human is not an oversight on a site making real money.

The uncomfortable part, for publishers

Everything above is about buyers avoiding MFA. There is a second story, and it is the one that matters if you run a real site.

The arbitrage arithmetic is genuinely tempting. Buy traffic at a $3 CPM, monetise at $10, keep $7 per thousand. That margin does not care whether the underlying site is a ghost or a twenty-year-old publication with a newsroom. A legitimate publisher who starts buying traffic to fill inventory is doing the same arithmetic as an MFA operator, with better content.

And it works, until it does not. The failure mode is not a sudden ban. It is gradual: the bought traffic engages worse than the organic audience, so the site's aggregate engagement metrics fall, so its inventory looks worse to buyers, so CPMs drop, so more bought traffic is needed to hit the same revenue. The publisher ends up with a larger, cheaper, worse audience and a brand that no longer commands what it used to.

The version of this that catches honest publishers

There is a specific trap worth naming, because it catches people with no intention of running an arbitrage operation.

A publisher has a good month, gets used to the revenue, then has a bad month. Traffic is down for reasons outside their control — an algorithm update, a seasonal dip, a story that did not land. Rather than accept the lower number, they buy some traffic to fill the gap. It works. The revenue line holds.

Next month they buy a bit more, because now that is the baseline. Within a year a meaningful share of the audience is purchased, the engagement average has fallen because bought traffic engages worse, and the CPMs have softened in response — which means more bought traffic is needed to hit the same number.

Nobody made a decision to become an arbitrage site. They made twelve small decisions to protect a monthly figure. That is how most of it happens, and it is why the honest advice is not "never buy traffic" but "know what percentage of your audience you bought, and watch that number like it is a debt" — because functionally it is one.

Where the line actually is

Three questions, and they are the same three a network should ask before representing anyone:

• Would this page exist if there were no ads on it? A real publication would still publish, just less profitably. An MFA page would have no reason to exist at all.
• Does the site have an audience, or only a media buying account? Organic search and direct traffic are the evidence. Their near-total absence is the diagnosis.
• Would you show the page to the advertiser? The screenshot test is unscientific and remarkably reliable. If the answer is no, the CPM is not the problem.

That judgement is exactly what a network has to make at the point of admission, and it is not something a filter can do for you — which is why partner vetting and traffic screening are the same discipline rather than two.

Both are described by what they do rather than by any result, because on this subject the only numbers worth quoting belong to the ANA and HUMAN, and they are above with their names attached.

What actually fixes this

Not a filter. The honest answer has three parts and none of them is a product you can buy on its own.

Buyers have to state a preference the machines can act on. Inclusion lists, supply-path optimisation, MFA-avoidance settings that are actually switched on. The tooling now exists; using it is a decision, and one that costs a little delivery efficiency in exchange for not funding the category.

Sell-side intermediaries have to judge at admission. Every MFA site is represented by somebody. Networks, SSPs and exchanges are the layer with both the visibility and the commercial relationship to say no, and the layer with the least incentive to, because the inventory converts. That tension is not going to resolve on its own.

Publishers have to be able to prove they are not it. This is the part that matters most to anyone reading this who runs a real site. As buyers get more suspicious of the long tail, the burden shifts to legitimate publishers to demonstrate an actual audience — organic share, returning visitors, session depth, a traffic history that predates the conversation. A site that cannot produce that evidence gets priced like the sites it cannot distinguish itself from.