The most talked-about targeting method of 2026 is older than the banner ad. Before anyone could follow a user across the web, advertisers bought the page. You put the watch ad in the watch magazine. You bought the back cover of the in-flight magazine because people on planes buy luggage. The first media plan I ever touched had no audiences in it at all, just a list of placements and a theory about who reads what. We called it media planning. The industry now calls it contextual targeting, and it is being rediscovered with the enthusiasm usually reserved for genuinely new things.
It is worth being precise about why it went away, because the reasons it died are the reasons it is back.
What killed it the first time
Contextual lost to behavioral for one reason: the cookie made people cheaper than pages. Why pay a premium to reach readers of a finance section when you could follow the same person to a recipe blog and reach them for a third of the CPM? The logic was airtight as long as two things stayed true. The identifier had to persist, and the measurement had to be believed.
There was also a self-inflicted wound. Early contextual was keyword matching, and keyword matching is dumb in ways that made for good conference slides. Airline ads next to plane crash coverage. Kitchen knife ads on crime stories. Brand safety tools built on the same crude matching then overcorrected, blocking the word "shot" on basketball coverage and defunding half the news industry by accident. Contextual earned a reputation as the clumsy option, and the reputation outlived the clumsiness.
The quiet collapse of the alternative
The behavioral machine did not fail loudly. It eroded. Apple's App Tracking Transparency took consent rates to the twenties and gutted mobile identifiers. Safari and Firefox, roughly four in ten browsing sessions depending on your category, have blocked third-party cookies for years. Chrome spent four years announcing deprecation, delaying it, and finally settling on user choice, which is deprecation with extra steps. Signal did not disappear on a date. It leaked out slowly enough that plenty of dashboards never admitted it.
What fills the gap is modeling. Modeled conversions, modeled audiences, modeled reach. Some of it is good modeling. All of it is a vendor grading their own homework, which is a subject I have written about in the number your CFO actually believes. The practical effect for a buyer is that a growing share of "audience" targeting is inference wearing the costume of observation. You are often paying a data fee for a guess.
Meanwhile the ceiling on identity-based tactics keeps dropping. Match rates govern everything downstream of a customer list, and match rates are the ceiling on everything else. When half your file cannot be matched, the most sophisticated audience strategy in the world runs on the half that could.
What actually changed in the machine
Here is the part that makes 2026 contextual a different product from 2012 contextual, and it is not marketing spin. The underlying text understanding moved from keyword lookup to language models. A modern contextual engine does not see the word "shot" and panic. It reads the page the way a person skims it: this is a basketball recap, the sentiment is celebratory, the entities are a team and a player, the category is sports, the commercial signal is athletic apparel and tickets, not trauma counseling.
Three specific upgrades matter. First, semantic classification at the page level, not the domain level. The finance section of a general news site is finance inventory now, not "news." Second, video understanding. Speech-to-text plus frame analysis means CTV and YouTube inventory can be bought against what is actually on screen, scene by scene, which was science fiction when contextual died the first time. Third, the plumbing got standardized. Seller-defined audiences and curation marketplaces let publishers package context with first-party signals in ways a buyer can actually transact on programmatically.
None of this requires knowing who the user is. That is the whole trick. The page does not need consent to be about something.
The economics nobody puts on the slide
Contextual inventory tends to clear cheaper than the identical impression wrapped in an audience segment, and the gap is usually the data fee plus the bid competition that identity attracts. On the plans I run, comparable placements typically price 20 to 40 percent below their behavioral twins before you count the segment fees, which run one to three dollars of CPM on their own. You are not paying a middleman to confirm the reader of a mortgage rate comparison page is thinking about mortgages.
The second economic advantage is reach that others structurally cannot buy. Roughly 40 percent of the open web is invisible to cookie-based targeting. That inventory is not worse. It is unpriced, because the demand that would price it cannot see it. Buying well there is one of the few remaining structural arbitrages in programmatic.
The third is legal gravity. A targeting method that never touches personal data does not care about the next state privacy law, the next consent framework, or the next platform policy shift. For regulated categories this is not a nice-to-have. Health advertisers who cannot use interest segments without a compliance review can buy condition-adjacent content all day, because targeting the article is not targeting the patient.
Where it wins, where it loses
Contextual is strongest where interest and content sit close together. Finance, autos, travel, B2B software, home improvement, anything with an active research phase. The person reading the comparison review is the market. It is also quietly excellent on CTV, where show-level and genre-level context plus a co-viewing audience beats a modeled household segment more often than the segment vendors would like you to know.
Be honest about where it loses. Contextual cannot retarget your cart abandoners; nothing about a page tells you someone left size 11 trail runners behind. It cannot frequency-cap a person across contexts, only a device in a session, so heavy contextual plans need creative rotation doing the fatigue work. And it cannot find the buying committee at a specific company; firmographic B2B still needs identity or a walled garden. The correct posture is not contextual instead of everything. It is contextual as the scalable base layer, identity spent where identity is provably worth its tax.
Measurement deserves one plain sentence: if your proof of contextual performance is click-based attribution, you have no proof. Page-level buys demand geo holdouts, matched-market tests, or a media mix model. The methods are older and slower and they are the ones a finance team will believe.
A worked example
A financial services client of mine ran the comparison properly last year, and the shape of the result is worth more than the confidentiality-safe numbers. Two line items, same creative, same flight, same DSP. The behavioral line bought a third-party "in-market: personal loans" segment at a $14 blended CPM, of which $2.40 was the segment fee. The contextual line bought pages classified around rate comparisons, debt consolidation, and credit education at a $9 CPM, no data fee. Click-based attribution said the behavioral line was winning by 30 percent, which surprised nobody, because segments built from browsing behavior harvest people already in motion and the click model hands them the credit.
Then we split geographies and held out markets, the way you do when you want an answer instead of a report. Measured on incremental applications, the contextual line produced a cost per incremental application 22 percent below the behavioral line. Not because the audience data was wrong about who was in-market. Because the contextual line reached in-market people the segment could not see, at a price that did not carry the identity tax, and the behavioral line was spending a meaningful share of its budget re-finding people the brand would have converted anyway. The dashboard and the truth disagreed, and the truth was cheaper.
One result is one result. I have also seen contextual lose a matched test in a category where purchase intent leaves no content trail. The point is not that context always wins. The point is that the only way to know is a test design that does not depend on the identifiers whose absence you are testing around.
How to run it like you mean it
Build the topic list from your own language data, not the vendor's taxonomy. Your search query reports and site search logs are a map of the phrases your buyers actually use; a contextual plan built from them beats a checkbox called "Business & Finance" by miles. Demand page-level transparency in reporting, because "contextual" from some vendors is still domain lists in a trench coat. Run it head-to-head against your behavioral line items in matched geos for a quarter and let the incrementality decide the mix. And pair it with creative built for the context, because when the page is the targeting, the creative is the conversion mechanism. An ad that acknowledges where it is running outperforms the same ad pretending it is everywhere.
There is a second-order opportunity hiding in the brand-safety wreckage: news. Years of keyword blocklists trained an entire industry to avoid the most-read, most-trusted, most attention-dense inventory on the web because a crude tool could not tell a war report from a recipe. Modern classification can. Advertisers who rebuilt news inclusion lists with sentiment-aware tools are buying premium attention at discount prices while their competitors still run 2019 blocklists. The data on this is unambiguous enough that avoiding all news coverage is now a performance decision, not a safety one, and it is the wrong performance decision.
The industry spent fifteen years treating context as the fallback for when data was missing. That had it backwards. Context was the original data. The pendulum is not swinging back to it out of nostalgia; the alternatives got expensive, legally fragile, and increasingly fictional, while the oldest signal in the business learned to read. The advertisers winning the next few years will be the ones who noticed first.