Google shipped a button that lets your readers put you inside AI answers. We downloaded it and weighed it before deciding whether to install it.

For two years the story of search has been the same story: an answer appears at the top of the page, the answer is assembled from your work, and the click that used to follow it does not. We have written about the regulator-forced opt-out, about what to publish so an assistant can quote you, and about how much of the web is now machine-written. Every one of those pieces ends in the same uncomfortable place. You can improve your odds. You cannot decide the outcome. An algorithm decides.

On August 20, 2026, Google published something with a genuinely different shape. It is a button. You put it on your own page. A reader clicks it, and that reader has now told Google to surface your site to them, by name, in Top Stories and inside AI Overviews and AI Mode. No algorithm adjudicates that choice. A person makes it.

That is unusual enough to be worth taking seriously. It is also unusual enough to be worth checking carefully, because a lever that publishers have wanted for two years is exactly the kind of thing that gets over-described in the first week. So we did both. We read the primary documentation, we reconciled the numbers against Google’s own posts, and then we downloaded the script Google asks you to embed and weighed it.

What was actually announced

Preferred Sources is not new. The button is.

Date	What shipped	Unique sources selected
June 26, 2025	Search Labs experiment, US and India, English	not disclosed
August 12, 2025	Out of Labs in Top Stories, US and India	not disclosed
December 10, 2025	Global rollout for English worldwide	nearly 90,000
April 30, 2026	Extended to all languages Google Search supports	not disclosed
May 27, 2026	Preferred sources labelled inside AI Overviews and AI Mode	more than 345,000
August 20, 2026	Embeddable button plus Search Central documentation	more than 600,000

The August 20 announcement was written by Mrinalini Loew, General Manager of the Google Search Ecosystem, and it contains one new capability plus two unrelated personalization features (natural-language topic tuning in Discover, and customizable audio briefings in Google News on Android). The new capability is this: publishers can now put the opt-in on their own page, instead of hoping readers find Google’s settings screen on their own.

That distinction is the whole product. Before August 20, adding a preferred source meant going to your Google account, finding Search personalization, finding Source preferences, and typing in a domain. Almost nobody does that for a site they like. After August 20, it is a button at the bottom of an article, and clicking it returns the reader to the exact spot they were reading.

What being a preferred source does, precisely

Google’s help page for readers is the clearest statement of the mechanism, and it is deliberately modest: “You’ll now find these sources more often in Top stories when they publish fresh content related to your search.”

More often. Not always, not first, not instead of. The publisher documentation adds that preferred content carries a visible “preferred” badge, and that the surfaces are Top Stories, available globally in every language where Google Search is available, plus AI Mode and AI Overviews in every language and locale where those two features exist.

John Mueller has been explicit that this runs alongside the ranking systems rather than overriding them, and that framing matches everything in the documentation. Nothing here changes what Google thinks of your site. It changes what one signed-in reader sees, for queries where you were already a candidate.

There is a second sentence on Google’s reader-facing help page that matters more to a small business than anything in the announcement, and we have not seen it quoted anywhere: “sources that aren’t updated regularly may not be available.”

That is the eligibility gate, stated plainly, and it is not about size. It is about cadence. A site that publishes is selectable. A five-page brochure site that has not changed since 2023 may simply not appear in the tool for a reader to pick, no matter how good it is.

The publisher documentation adds one hard structural rule: only domain-level and subdomain-level sites are eligible to appear in the source preferences tool. https://www.example.com/ works, https://code.example.com/ works, https://www.example.com/blog does not. If you run several distinct publications under one domain as subdirectories, readers cannot pick one of them. They pick all of you or none of you.

The three ways to install it

Google documents three integrations, and the difference between them is the entire practical decision.

Option one, the recommended two-liner. A script tag and a div:

<script async src="https://news.google.com/swg/js/v1/publisher.js"></script>
<div google-add-preferred-source-btn></div>

That renders a Google-styled, automatically localized button. It accepts data-theme (light or dark) and data-lang to override the language.

Option two, the advanced integration, for teams that want their own button design. Either the ES module:

import { preferredSource } from "https://news.google.com/swg/js/v1/publisher.mjs";

preferredSource.init({ theme: 'light', lang: 'en' });

document.querySelector('#myButton').onclick = () => {
  preferredSource.addPreferredSource();
};

or the classic-script callback queue, using preferred-sources-control="manual" on the script tag and pushing a function onto self.PREFERRED_SOURCE.

Option three, the deeplink. No JavaScript at all:

<a href="https://www.google.com/preferences/source?q=example.com">Add as Preferred Source</a>

That sends the reader to Google’s source preferences tool with your domain pre-filled. They confirm there rather than on your page, and they do not automatically bounce back to what they were reading. More friction, and no third-party code.

Worth reading the documentation’s own closing note before you pick, because it defuses the urgency: “These methods are examples of how you can build your audience and help people find your site as a preferred source. It’s not required to do them in order to appear as a preferred source.” You are already selectable. The button only shortens the path.

We weighed the script

Nobody writing about this appears to have opened publisher.js. We fetched it from this machine on the morning of September 6, 2026, the same way a browser would, and then read it.

Measurement	Value
Uncompressed size	138,068 bytes
Transferred size (brotli/gzip)	38,455 bytes
cache-control	public, max-age=3000
last-modified	Wed, 02 Sep 2026 20:15:29 GMT
access-control-allow-origin	*

Three things follow, and each of them is a cost somebody has to decide is worth paying.

It is not a button library. It is the Subscribe with Google SDK. The swg in the path is the giveaway, and the bundle confirms it. The iframe it injects into your page is created with the title Subscribe with Google Service. The bundle carries the constants for an entire monetization product: TYPE_SUBSCRIPTION, TYPE_CONTRIBUTION, TYPE_REGISTRATION_WALL, TYPE_NEWSLETTER_SIGNUP, TYPE_REWARDED_SURVEY, TYPE_BYO_CTA, TYPE_REWARDED_AD. To render one button, you load the paywall, contribution, registration-wall and rewarded-ad machinery too. None of it runs for the preferred-sources flow. All of it parses.

It caches for fifty minutes. max-age=3000 is an unusual choice for a versioned library path. Compare it to the immutable, one-year cache we put on our own hashed assets. A reader who visits you twice in a week re-fetches this file both times. That is not fatal, it is just the difference between a 38 KB one-time cost and a 38 KB recurring one, and it is the sort of thing that quietly shows up in a Lighthouse run later.

It reads your page context. The library passes window.location.href and document.referrer into the service, and on button click it logs an event (ACTION_ADD_PREFERRED_SOURCES_BUTTON_CLICK) with the page’s canonical URL attached. That is not a scandal, it is roughly what any embedded Google widget does, and it is what you are agreeing to. Worth knowing before you agree to it. In the library’s favour: there are no references to document.cookie or localStorage anywhere in the bundle. Whatever storage happens, happens inside Google’s frame on Google’s origin, not in yours.

The frames themselves come from a constant in the bundle, frontEnd: "https://news.google.com", with URLs built under /swg/ui/v1/. The library supports three transport modes: iframe, popup, and redirect.

What it costs on a site with a real CSP

This is where the two-liner stops being a two-liner. Our Content-Security-Policy starts at default-src 'self', which means Google’s script is blocked, and so is Google’s frame.

To ship option one on this site we would need, in vercel.json:

- script-src 'self' 'unsafe-inline' https://va.vercel-scripts.com;
+ script-src 'self' 'unsafe-inline' https://va.vercel-scripts.com https://news.google.com;
+ frame-src https://news.google.com;

The frame-src line is the one people miss. We do not currently declare frame-src at all, so it inherits from default-src 'self', and the injected iframe fails silently with a console error and a button that does nothing.

And because we also emit a per-page meta CSP through Astro’s security.csp, the same origin has to be added a second time, in astro.config.mjs, to scriptDirective.resources and as a new frame-src directive. Two files, or the change half-works in a way that is annoying to debug. If you run a strict CSP and a build-time CSP, check both.

One more interaction that is easy to miss: we send Cross-Origin-Opener-Policy: same-origin. If the library ever falls back to its popup mode, COOP severs the opener relationship with a cross-origin popup, so the popup cannot talk back to the page that opened it. The iframe path is unaffected. If you serve COOP and you install this, test the flow rather than assuming it.

Our decision for this site: we are not loading Google’s script. The honest arithmetic is that we would be adding 38 KB over the wire, a third-party frame origin, and two CSP relaxations, in exchange for a slightly smoother version of an opt-in that the deeplink already provides for zero bytes. For a news operation running Subscribe with Google anyway, where the library is already on the page, option one is free and obviously correct. For a business site with a hardened header, the deeplink is the better trade, and it is the version we will put in front of clients.

What the coverage is getting wrong

We checked four claims that are being repeated widely. Three of them are wrong or unsupported.

The number. TechCrunch’s August 20 article reports “over 345,000 unique sources selected through this feature since May 2026 launch.” Google’s own post the same day says “more than 600,000 unique sources.” The 345,000 figure is the May 27 number, and it was already three months stale on the day it was published. The gap between the two, roughly 255,000 selections in under three months, is the actual growth story and it is more interesting than either number alone.

The date. Several write-ups describe a May 2026 launch. Preferred Sources launched out of Labs on August 12, 2025, went global on December 10, 2025, and reached all supported languages on April 30, 2026. May 27, 2026 was the fourth milestone, not the first: it is when preferred sources started being labelled inside AI Overviews and AI Mode. That correction matters because it changes how you read the adoption curve. This is not a feature that found 600,000 sources in three months. It took thirteen.

Discover. A lot of coverage says the button boosts you in Search, Discover and Google News. Google’s post is titled “Personalize the content you see on Search, Discover, and News”, and it does announce Discover and News features, but they are the separate topic-tuning and audio-briefing features. The publisher documentation, twice, names exactly three surfaces for preferred sources: Top Stories, AI Overviews, AI Mode. If you are budgeting effort against an expected Discover lift, you are budgeting against something Google has not claimed.

The one that is true but soft: “twice as likely to click.” This claim is real and it is Google’s. Nick Fox, SVP of Knowledge and Information, first stated it on December 10, 2025, and it has been repeated in May and August 2026. What has never been disclosed is the baseline. Twice as likely compared to what: the same site’s unpreferred impressions, the average result on the page, a matched query set? Nor whether the effect holds over time. And the population is self-selected in the most direct way possible. Somebody who deliberately types your domain into a settings screen was always going to click you more than a stranger. Some of that doubling is the feature working. Some of it is the kind of person who uses the feature. Google has not separated the two, and until it does, treat the number as directional rather than as a forecast.

So is this worth anything to a small business?

Here is the part that decides it, and it cuts against most of the advice being written this week.

Preferred Sources is a retention lever, not an acquisition lever. It does nothing at all for a person who has never heard of you. It only affects the results of someone who already read you, liked you enough to click, and is signed in to a Google account. Its value scales with the size of the audience you already have.

That is precisely backwards from what most small businesses need. If you are a Toronto plumbing company, your problem is not that your existing readers forget you. It is that people searching for a plumber have never seen your name. This button does not touch that problem. Neither, for that matter, will Top Stories: that surface is for news queries, and your service pages are not going to appear in it regardless of how many people prefer you.

So the sober assessment for most local businesses is: install the deeplink because it costs you one line of HTML, and expect nothing measurable from it this quarter. Then go back to the work that actually generates discovery, which for a Toronto business is still the map pack and your Business Profile.

The businesses where this genuinely matters are the ones that publish. If you run a clinic that posts patient guidance monthly, a trades company with a real project blog, a SaaS product with a changelog people read, or a firm whose newsletter has a list behind it, then you have exactly the asset this feature rewards: repeat readers who would say yes if asked. Ask them. It is the same motion as a newsletter signup, with less commitment from the reader and no list for you to maintain.

And there is a second-order reason to take it seriously even at zero measurable lift. AI Overviews and AI Mode decide, per query, which handful of sources get cited. A per-user preference that survives across sessions is one of the very few inputs to that decision that is not an algorithm’s opinion of you. Cheap options on that kind of thing are worth holding.

Do this, in order

1. Check whether you are selectable. Signed in to a Google account, open https://www.google.com/preferences/source?q=yourdomain.com. If your site appears, readers can pick you. We tried this signed out while researching this piece and got back a shell page with a sign-in controller and no source list, so this check genuinely requires an account. There is no signed-out way to verify it.

2. If you are not there, fix cadence before anything else. Google’s own wording is that sources which are not updated regularly may not be available. No button helps a site that cannot be selected.

3. Pick the deeplink unless you already run Subscribe with Google. One line, no script, no CSP change, no third-party frame:

<a href="https://www.google.com/preferences/source?q=yourdomain.com"
   rel="nofollow">Add us as a preferred source on Google</a>

4. If you do install the script, budget the real cost. 38 KB over the wire, a fifty-minute cache, one new script origin and one new frame origin in your CSP, in every place you declare a CSP. Then load a page and confirm the button renders rather than assuming it does.

5. Put it where a satisfied reader is, not where a new one is. End of article, end of a guide, inside your newsletter confirmation page. A modal on a first visit will earn you nothing except a worse INP score.

6. Measure the only thing you can measure, and label it honestly. There is no Preferred Sources report in Search Console. You can instrument clicks on your own link or button, and that is all. A click is not a selection: the reader still has to confirm on Google’s side, and nothing tells you whether they did. Anyone selling you a preferred-sources conversion rate is selling you a click-through rate on a link.

7. Do not expect ranking movement, and do not let anyone bill you for it. This is per-user personalization sitting alongside ranking. If your rankings change next month, something else changed them.

The part worth sitting with

The interesting thing about this button is not the traffic. It is what it says about where distribution is going.

Google spent twenty years insisting that relevance was a computed property of the web and that the answer to “why am I not ranking” was always “be better.” Preferred Sources is a quiet admission that in an AI-answer interface, computed relevance is not sufficient, because when the interface returns one synthesized paragraph, somebody has to decide whose sentences are in it, and the algorithmic version of that decision is turning out to be unpopular with everyone including regulators.

So Google is doing what platforms do when the algorithm becomes the story: handing a slice of the decision back to the user, and asking publishers to collect the opt-ins on their own pages, using their own real estate, at their own performance cost. You are being asked to help build Google’s personalization graph. That is a fair trade only if you get something back, and right now the something is unmeasurable by design, because the reporting does not exist.

Take the deal anyway, in its cheapest form. One line of HTML, aimed at the readers who already like you, is not a strategy and it is not a risk. Just do not confuse it with the work. The thing that gets a stranger to your site in 2026 is still the same thing it was in 2019: being genuinely the best answer to a question somebody is actually asking, published somewhere a machine can read it.

If you want us to look at whether your site is selectable, whether your publishing cadence clears Google’s bar, and what a preferred-source ask would cost you in page weight, send us the URL. It is a ten minute check and it does not cost anything.

Sources

• Personalize the content you see on Search, Discover, and News, Mrinalini Loew, General Manager, Google Search Ecosystem, August 20, 2026, retrieved September 6, 2026, for the button announcement and the 600,000 unique sources figure
• Guide to Preferred Sources in Google Search for Web Publishers, Google Search Central, last updated August 20, 2026, retrieved September 6, 2026, for all three integration options, the domain and subdomain eligibility rule, the supported surfaces and the note that the button is not required
• Preferred Sources in Google Search, Google Search Help, retrieved September 6, 2026, for the reader-facing behaviour, the settings sync statement and the wording that sources which are not updated regularly may not be available
• New ways to find your favorite sources and original content in AI Search, Duncan Osborn, Product Manager, Google Search, May 27, 2026, retrieved September 6, 2026, for the AI Overviews and AI Mode expansion, the 345,000 figure and the twice-as-likely claim
• How to select your preferred sources in Top Stories in Search, Duncan Osborn, August 12, 2025, retrieved September 6, 2026, for the original US and India launch
• Google Preferred Sources gains 255,000 selections since late May, PPC Land, August 2026, retrieved September 6, 2026, for the consolidated timeline and the observation that no downstream measurement exists in Search Console
• Google’s Preferred Sources goes global with doubled click rates, PPC Land, December 2025, retrieved September 6, 2026, for the December 10, 2025 global rollout, the June 26, 2025 Labs start and the Nick Fox click claim
• Google rolling out Preferred Sources globally and announces Spotlighting subscriptions, Search Engine Land, December 10, 2025, retrieved September 6, 2026, for the nearly 90,000 unique sources figure and the English-worldwide-first rollout
• Google expands ‘Preferred Sources’ to everyone so you actually see your favorite sites, 9to5Google, April 30, 2026, retrieved September 6, 2026, for the expansion to all supported Search languages
• Google gives publishers a new way to fight AI-driven traffic losses, TechCrunch, August 20, 2026, retrieved September 6, 2026, the source of the 345,000 figure discussed above
• Google AI Overviews and AI Mode gain preferred sources, plus new perspectives carousel and highly cited labels, Search Engine Land, May 27, 2026, retrieved September 6, 2026
• Google Preferred Sources Hit 345K, Expand Into AI Search, Search Engine Journal, May 27, 2026, retrieved September 6, 2026, for John Mueller’s clarification that the feature works alongside ranking systems
• https://news.google.com/swg/js/v1/publisher.js, fetched and analysed September 6, 2026, for the file sizes, response headers, Subscribe with Google constants, iframe origin and event logging described above