With the upcoming removal of third-party cookies on the web, key use cases that browsers want to support will need to be addressed with new APIs. One of those use cases is interest-based advertising.
Interest-based advertising (IBA) is a form of personalized advertising in which an ad is selected for the user based on interests derived from the sites that they’ve visited in the past. This is different from contextual advertising, which is based solely on the interests derived from the current site being viewed (and advertised on). One of IBA’s benefits is that it allows sites that are useful to the user, but perhaps could not be easily monetized via contextual advertising, to display more relevant ads to the user than they otherwise could, helping to fund the sites that the user visits.
The intent of the Topics API is to provide callers (including third-party ad-tech or advertising providers on the page that run script) with coarse-grained advertising topics that the page visitor might currently be interested in. These topics will supplement the contextual signals from the current page and can be combined to help find an appropriate advertisement for the visitor.
Example usage to fetch an interest-based ad:
// document.browsingTopics() returns an array of up to three topic objects in random order.
const topics = await document.browsingTopics();
// The returned array looks like: [{'value': Number, 'taxonomyVersion': String, 'modelVersion': String}]
// Get data for an ad creative.
const response = await fetch('https://ads.example/get-creative', {
method: 'POST',
headers: {
'Content-Type': 'application/json',
},
body: JSON.stringify(topics)
});
// Get the JSON from the response.
const creative = await response.json();
// Display ad.
The topics are selected from an advertising taxonomy. The initial taxonomy (proposed for experimentation) will include somewhere between a few hundred and a few thousand topics (our initial design includes ~350 topics; as a point of reference the IAB Audience Taxonomy contains ~1,500) and will attempt to exclude sensitive topics (we’re planning to engage with external partners to help define this). The eventual goal is for the taxonomy to be sourced from an external party that incorporates feedback and ideas from across the industry.
The topics will be inferred by the browser. The browser will leverage a classifier model to map site hostnames to topics. The classifier weights will be public, perhaps built by an external partner, and will improve over time. It may make sense for sites to provide their own topics (e.g., via meta tags, headers, or JavaScript) but that remains an open question discussed later.
- Note that this is an explainer, the first step in the standardization process. The API is not finalized. The parameters below (e.g., taxonomy size, number of topics calculated per week, the number of topics returned per call, etc.) are subject to change as we incorporate ecosystem feedback and iterate on the API.
document.browsingTopics()
returns an array of up to three topics, one from each of the preceding three epochs (weeks). The returned array is in random order.- By providing three topics, infrequently visited sites will have enough topics to find relevant ads, but sites visited weekly will learn at most one new topic per week.
- The returned topics each have a topic id (a number that can be looked up in the published taxonomy), a taxonomy version, and a classifier version. The classifier is what maps the hostnames that the user has visited to topics.
- The topic id is a number as that’s what is most convenient for developers. When presenting to users, it is suggested that the actual string of the topic (translated in the local language) is presented for clarity.
- The array may have zero to three topics in it. As noted below, the caller only receives topics it has observed the user visit in the past.
- For each week, the user’s top 5 topics are calculated using browsing information local to the browser. One additional topic, chosen uniformly at random, is appended for a total of 6 topics associated with the user for that week/epoch.
- When
document.browsingTopics()
is called, the topic for each week is chosen from the 6 available topics as follows:- There is a 5% chance that the random topic is returned
- Otherwise, return the real top topic whose index is HMAC(per_user_private_key, epoch_number, top_frame_registrable_domain) % 5
- Whatever topic is returned, will continue to be returned for any caller on that site for the remainder of the three weeks.
- The 5% noise is introduced to ensure that each topic has a minimum number of members (k-anonymity) as well as to provide some amount of plausible deniability.
- The reason that each site gets one of several topics is to ensure that different sites often get different topics, making it harder for sites to cross-correlate the same user.
- e.g., site A might see topic ‘cats’ for the user, but site B might see topic ‘automobiles’. It’s difficult for the two to determine that they’re looking at the same user.
- The beginning of a week is user specific and may include some noise. That is, not everyone will calculate new topics at the same time on the same day.
- When
- Not every API caller will receive a topic. Only callers that observed the user visit a site about the topic in question within the past three weeks can receive the topic. If the caller (specifically the site of the calling context) did not call the API in the past for that user on a site about that topic, then the topic will not be included in the array returned by the API.
- This is to prevent the direct dissemination of user information to more parties than the technology that the API is replacing (third-party cookies).
- Example:
-
Week 1: The user visits a bunch of sites about fruits, and the Topics taxonomy includes each type of fruit.
Site Visited Site Topic(s) Parties that call the Topics API on the site A Apples, Grapes T, R B Bananas S C Cantaloupe T, S D Durian E Emblica S F Apples S G Grapes S -
End of week 1: The Topics API generates the user’s top topics for the week, and which third parties are allowed to learn about them:
Top Topic Parties That Can Learn About the Topic Apples T, R, S Bananas S Cantaloupe T, S Emblica S Grapes T, R, S -
In week 2, if a caller on any site calls the API, then the returned topic array will only include topics for which the caller is in the “Parties that Can Learn About the Topic” column for that topic for that site for that epoch.
-
If, in week 2, T is added to site B, and sees that the user is interested in Bananas, they won’t be able to receive the Bananas topic until the following epoch, in week 3. This ensures that third-parties don’t learn more about a user’s past (e.g., past interest in Bananas) than they could with cookies.
-
- The history window included in the calculation of topics available to each caller is three weeks.
- Only topics of sites that use the API will contribute to the weekly calculation of topics.
- Further, only sites that were navigated to via user gesture are included (as opposed to a redirect, for example).
- If the API cannot be used (e.g., disabled by the user or a response header), then the page visit will not contribute to the weekly calculation.
- Interests are derived from a list or model that maps website hostnames to topics.
- The model may return zero topics, or it may return one or several. There is not a limit, though the expectation is 1-3.
- We propose choosing topics of interest based only on website hostnames, rather than additional information like the full URL or contents of visited websites.
- For example, “tennis.example.com” might have a tennis topic whereas example.com/tennis would only have topics related to the more general example.com.
- This is a difficult trade-off: topics based on more specific browsing activity might be more useful in picking relevant ads, but also might unintentionally pick up data with heightened privacy expectations.
- Initial classifiers will be trained by Google, where the training data is human curated hostnames and topics. The model will be freely available (as it is distributed with Chrome).
- It’s possible that a site maps to no topics and doesn’t add to the user’s topic history. Or it’s possible that a site adds several.
- The mapping of sites to topics is not a secret, and can be called by others just as Chrome does. It would be good if a site could learn what its topics are as well via some external tooling (e.g., dev tools).
- The mapping of sites to topics will not always be accurate. The training data is imperfect (created by humans) and the resulting classifier will be as well. The intention is for the labeling to be good enough to provide value to publishers and advertisers, with iterative improvements over time.
- Please see below for discussion on allowing sites to set their own topics.
- The mapping of hostnames to topics will be updated over time, but the cadence is TBD.
- How the top five topics for the epoch are selected:
- At the end of an epoch, the browser calculates the list of eligible pages visited by the user in the previous week
- Eligible visits are those that used the API, and the API wasn’t disabled
- e.g., disabled by preference or site response header
- Eligible visits are those that used the API, and the API wasn’t disabled
- For each page, the host portion of the URL is mapped to its list of topics
- The topics are accumulated
- The top five topics are chosen
- At the end of an epoch, the browser calculates the list of eligible pages visited by the user in the previous week
- If the user opts out of the Topics API, or is in incognito mode, or the user has cleared all of their history, the list of topics returned will be empty
- We considered a random response instead of empty but prefer empty because:
- It’s clearer to users to see that after disabling the feature, or entering incognito, that no topic is sent
- Returning random values would be a loss to utility, for marginal gain in privacy, since the API will return an empty topic for one of many reasons:
- The user is in incognito mode
- The caller hasn’t seen the topic
- The user's browsing history has been cleared
- The API is disabled
- We considered a random response instead of empty but prefer empty because:
- If the user doesn’t have enough browsing history to create 5 topics:
- Note that this is still being decided
- Depending on how many topics the user does have, the remaining may be randomly generated. Else no topics will be created for that epoch.
- A site can forbid topic calculation on a page via the following permission policy header:
Permissions-Policy: browsing-topics=()
- Note: The old
Permissions-Policy: interest-cohort=()
from FLoC will also forbid topic calculation.
The Topics API’s mission is to take a step forward in user privacy, while still providing enough relevant information to advertisers that websites can continue to thrive, but without the need for invasive tracking enabled via existing tracking methods.
What does it mean to take a step forward in privacy for Interest Based Advertising?
- It must be difficult to reidentify significant numbers of users across sites using just the API.
- The API should provide a subset of the capabilities of third-party cookies.
- The topics revealed by the API should be less personally sensitive about a user than what could be derived using today’s tracking methods.
- Users should be able to understand the API, recognize what is being communicated about them, and have clear controls. This is largely a UX responsibility but it does require that the API be designed in a way such that the UX is feasible.
We expect that this proposal will evolve over time, but below we outline our initial interpretation of how we think this proposal relates to our stated privacy goals.
- It must be difficult to reidentify significant numbers of users across sites using the API alone.
- In order to meet this goal, it is important that only a trace amount of information that could be used for cross-site reidentification is returned. In other words, the data must be rate-limited. The Topics API uses three mechanisms to slow the rate of leakage:
- Different sites will receive distinct topics for the same user in the same week. Since someone’s topic on site A usually won’t match their topic on site B, it becomes harder to determine that they’re the same user.
- The topics are updated on a weekly basis, which limits the rate of information dissemination.
- And finally, some fraction of the time, a random topic will be returned for a given site for that week.
- In order to meet this goal, it is important that only a trace amount of information that could be used for cross-site reidentification is returned. In other words, the data must be rate-limited. The Topics API uses three mechanisms to slow the rate of leakage:
- The API must not only significantly reduce the amount of information provided in comparison to cookies, it would also be better to ensure that it doesn’t reveal the information to more interested parties than third-party cookies would.
- In order to be a privacy improvement over third party cookies, the Topics API caller should learn no more than it could have using third-party cookies. This means the API shouldn’t inform callers about topics that the caller couldn’t have learned for itself using cookies. The topics that a caller could have learned about using cookies, are the topics of the pages that the caller was present on with that same user. This is why the Topics API restricts learning about topics to those callers that have observed the user on pages about those topics.
- Note that this means that callers with more third-party presence on sites the user visited will be more likely to have topics returned by
document.browsingTopics()
. - Also note that this implies that the API will generally be invoked from inside a third-party iframe, rather than inside the main frame of a page.
- The topics revealed by the API should be significantly less personally sensitive for a user than what could be derived using existing tracking methods.
- Third party cookies can be used to track anything about a user, from the exact urls they visited, to the precise page content on those pages. This could include limitless sensitive material. The Topics API, on the other hand, is restricted to a human curated taxonomy of topics. That’s not to say that other things couldn’t be statistically correlated with the topics in that taxonomy. That is possible. But when comparing the two, Topics seems like a clear improvement over cookies.
- Users should be able to understand the API, recognize what is being said about them, know when it’s in use, and be able to enable or disable it.
- By leveraging a human-readable taxonomy of topics, people can learn what is being said about them. They can remove topics they specifically don’t wish to see ads for, and there can be UX for informing the user about the API and how to enable or disable it. Further, topics are cleared when history is cleared.
FLoC ended its experiment in July of 2021. We’ve received valuable feedback from the community and integrated it into the Topics API design. A highlight of the changes, and why they were made, are listed below:
- FLoC didn’t actually use Federated learning, so why was it named Federated Learning of Cohorts?
- This is true. The intent had been to integrate federated learning into FLOC but we found that on-device computation offered enough utility and better privacy.
- FLoC added too much fingerprinting data to the ecosystem
- The Topics API significantly reduces the amount of cross-site identifiable information. The coarseness of the topics makes each topic a very weak signal; different sites receiving different topics further dilutes its utility for fingerprinting.
- Stakeholders wanted the API to provide more user transparency
- The Topics API uses a human-readable taxonomy which allows users to recognize which topics are being sent (e.g., in UX).
- Stakeholders wanted the API to provide more user controls
- With a topic taxonomy, browsers can offer a way (though browser UX may vary) for users to control which topics they want to include
- The Topics API will have a user opt-out mechanism
- FLoC cohorts might be sensitive
- FLoC cohorts had unknown meaning. The Topics API, unlike FLoC, exposes a curated list of topics that are chosen to avoid sensitive topics. It may be possible that topics, or groups of topics, are statistically correlatable with sensitive categories. This is not ideal, but it’s a statistical inference and considerably less than what can be learned from cookies (e.g., cross-site user identifier and full-context of the visited sites which includes the full url and the contents of the pages).
- FLoC shouldn’t automatically include browsing activity from sites with ads on them (as FLoC did in its initial experiment)
- To be eligible for generating users’ topics, sites wil have to use the API.
We consider the API to be a step toward improved user privacy on the web. It is, however, not perfect:
- It is possible for an entity (or entities) to cooperate across hosts and acquire up to 15 topics per epoch for the same user in the first week.
- If the user is known to be the same across colluding sites (e.g., because they’re logged into each with a persistent identifier), then it is possible for those sites to join their topics for the user together. This could also be achieved via adding topics to URLs when navigating between cooperating sites.
- Where does 15 come from? If the site has never seen the user before, then all 3 topics that it receives (from the past 3 epochs) will be new to it. And if the topics completely change for the user between epochs, then up to 15 of them will be new. it is unlikely that all 15 would be unique, but is still theoretically possible.
- In order for these topics to be revealed, they have to come from past observation from the given caller. Meaning that this is still a subset of the vast array of data that third-party cookies could divulge.
- A potential mitigation is to detect cases where a user has provided a site with a cross-site identifier (e.g., logged in) and in that case restrict the site to a particular index of the top 5 topics. So all such sites have the same topic choice, and less information is learned.
- Can anything other than topics be learned about the user’s browsing history?
- If a caller is only present on one site about a topic, and the call returns that topic, then the caller can infer the site that was visited.
- It is theoretically possible to have a number of different callers that call the API on different sets of sites collude to determine more detail about the sites a user visited, or to accumulate a user identifier over time. This is something that the browser could potentially observe and may intervene on if necessary.
- There is one piece of information that the API reveals that goes beyond the capabilities of third-party cookies: that the topic returned is one of the top 5 browsing topics for the user for the given week.
- The caller must have already known that the user visited a page about that topic in the past few weeks, but they didn’t necessarily know that it was one of the most frequent topics.
- We could alternatively allow each caller to have its own set of topics for a given user, which would prevent this leak. But it would allow a site to learn topics much faster if the various callers on the site communicate their topics with each other.
- Another possible mitigation is to pick the 5 topics at random, but weighted such that more frequently visited topics are more likely to be picked. This makes it a probabilistic determination that the topic was one of the top for the user for the week.
- If a caller is only present on one site about a topic, and the call returns that topic, then the caller can infer the site that was visited.
- There are means by which sensitive information may be revealed:
- As a site calls the API for the same user on the same site over time, they will develop a list of topics that are relevant to that user. That list of topics may have unintended correlations to sensitive topics.
- In the end, what can be learned from these human curated topics derived from the hostnames of pages that the user visits is probabilistic, and far less detailed than what cookies can provide from full page content, full urls, and precise cross-site identifiers. While imperfect, this is clearly better for user privacy than cookies.
This proposal benefited greatly from feedback from the community, and there are many ways to provide feedback on this proposal and participate in ongoing discussions, including responding on the linked issue in the repository or in a W3C group such as the Improving Web Advertising Business Group or the Private Advertising Technology Community Group. Some issues that we’d like to discuss:
- Should sites be able to set their topics, or should topics be determined by the browser or some third-party entity?
- If the client does it, where does the ML model come from? What data is it trained on?
- If the sites do it, might they pollute the algorithm by setting the topic to the most valuable?
- What should happen if a site disagrees with the topics assigned to it by the browser?
- Should there be a way to alter the assignment?
- Does mislabeling cause harm?
- What topic taxonomy should be used? Who should create and maintain it? How many topics should the taxonomy contain?
- Eventually it would be good if this was produced externally to the browser and became an industry standard.
- The taxonomy should be publicly available for transparency.
- If the number of topics increase, we’ll need to balance that with the ability of sites to observe topics (e.g., if there are more topics, there is less of a chance that an ad-tech has seen the chosen topic in the past).
- What standard might be used for determining which topics are sensitive?
- Should they be regional?
- How might the browser detect abusive usage of the API to keep the topic dissemination rate in line with expectations?
- Should sites receive historic topics every visit, or first visit only?
- For sites that users frequently visit there is no difference in privacy. For infrequently visited sites, this becomes a trade-off between topic dissemination rate and utility.
- How might one define “first visit”?
- It could be: does the site have any cookies or other storage for the user? If so, it’s not first visit.
- Should there be a way to send topics via Fetch as a request header?
- This would reduce the need for expensive (and slow) x-origin iframes to be created.