The Parity Benchmark Hiding in Google's Help Center
Google publicly documents seven contextual inputs to Ad Rank. The open web runs on roughly two of them. That page is the cleanest measure of the gap.
Nate Woodman ยท July 24, 2026
The best description of what the open web is missing is not in a leaked deck or a conference keynote. It is a public page in Google's own help center, written to educate advertisers, titled "About Ad Rank." Read it carefully and it functions as something Google never intended: a parity benchmark for everyone outside the walls.
What Google actually says about its own ranker
The page lays out six factors behind Ad Rank: your bid, the quality of your ads and landing page, rank thresholds, auction competitiveness, the expected impact of ad assets and formats, and the context of the person's search. It notes that Ad Rank is calculated every time a user searches and recalculated for different positions on the page.
The context factor is the one worth quoting. Google enumerates what it checks: the search terms the person entered, the person's location at the time of the search, the type of device, the time of the search, the nature of the search terms, the other ads and results on the page, and other user signals and attributes.
Count them. Google itself names seven distinct contextual inputs to its ranker, recalculated per impression. Context, to the operator of the world's largest ad system, is not one dimension. It is a suite.
The open web runs on two of the seven
Now hold the classical open-web bid request up against that list. What travels today is, roughly, an audience representation (segment IDs) and some page-level context (a URL, a category). Location arrives coarsely if at all, device as a crude type flag, time as a timestamp nobody's model reasons about, content adjacency almost never, and recent user response signal not at all outside each DSP's private silo.
Google publicly documents seven contextual inputs to Ad Rank. The open web has had access to roughly two of them, and it has spent fifteen years compensating with ever-more-elaborate audience taxonomies. The walled-garden magic was never a single secret model. It is a publicly described multi-input ranker, fed by inputs the open web's plumbing simply does not carry.
Why this page settles two arguments
The first argument it settles is whether multi-dimensional ranking is over-engineering. The standing objection to richer bid-stream context runs: does the open web not just need better audiences? Google's own help center answers no. Audiences are one input among seven, and the system recalculates the blend on every single search. When the operator of the most profitable ad platform in history splits context into seven enumerated inputs, a bid request that collapses audience and context into one field is not simpler. It is a generation behind.
The second argument it settles is whether parity work is speculative. The emerging open-web standards for carrying context as embeddings, user state, physical setting, time, media adjacency, creative semantics, each as its own first-class signal, are sometimes criticized as theoretical architecture. The criticism has to contend with the fact that the input list in those specs maps almost one-to-one onto what Google has openly documented about its own auction for years. Search terms map to the audience query, location and device to physical context, time of search to temporal context, surrounding results to media context, other user signals to identity and reinforcement state, ad quality to creative semantics. This is not invention. It is transcription.
The non-hostile source is the strongest kind
What makes this page so useful is precisely that it is not adversarial. Nobody leaked it. It is Google explaining itself to its own customers, in plain language, with no reason to overstate. That turns the open web's parity argument from speculation into parity-on-the-record: the target system's owner has described the target, publicly, in enough detail to build against.
It also quietly validates the direction of travel. The page's mention of expected impact from ad formats points at placement-level signal the standards bodies are still catching up to, and Google's other publications on YouTube recommendations document cohort and real-time event signals beyond even the seven on this page. The benchmark keeps moving. But that is the nature of benchmarks, and it is a far better position than the one the open web has occupied for a decade: competing against a system it told itself was unknowable, when the description was sitting in the help center the whole time.
The practical takeaway for anyone building or buying on the open web is one sentence. Stop asking whether multi-input ranking is coming, the incumbent has documented it for years, and start asking which of the seven inputs your stack can actually deliver.