Google's Spark Agent Never Stops Searching: How to Make Your Site Citable by 24/7 AI Agents

Google launched Gemini 3.7 Flash at half the price and put its personal Spark agent on top of it in more than 160 countries: an agent that searches, reads, verifies, and acts around the clock while the user isn't watching. For GEO this is a structural shift: the "user" reading your content is no longer always a person. Agents don't click, they cite, and when they cite, the human may never visit your site. Visibility now goes to the page the agent verifies and decides to quote, not the one with the best title. Here's what changed and what to do about it this week.

What Google announced on August 13

On August 13 Google introduced Gemini 3.7 Flash, which it calls its most intelligent workhorse model, with gains in software engineering, knowledge work, and web development workflows. The launch came just three weeks after the previous release, as Ars Technica noted, at an introductory rate of $0.75 per million input tokens and $3.75 per million output tokens, half of what 3.6 Flash cost.

The part that matters here isn't the model, it's what runs on it: Spark. Google showed Spark at I/O as the personal agent that works 24/7, and since the announcement it runs on Gemini 3.7 Flash for Google AI Pro and Ultra subscribers in more than 160 countries. One detail almost nobody mentioned: the footnote excludes the European Economic Area, the UK, Switzerland, and Nigeria. Latin America is not on that list.

Why the price matters more than the model

Every time inference gets cheaper, agents do more work. That's not a metaphor: an agent that could read 10 pages per task can now read 30 for the same cost. Verification gets deeper, scheduled tasks get more frequent, and the agent can afford to compare sources before citing.

Half price through the end of 2026 isn't a promotion, it's a signal that Google wants Spark to be the default search channel for its subscribers. More agent hours means more pages read, more claims verified, more summaries generated. Every summary ends in an answer with citations, and citations are the only traffic you'll get if the user never clicks.

The searcher changed: human vs agent

Optimizing for a person and optimizing for an agent aren't subtle variations, they're different mechanics:

• A human looks at the title and description, clicks two or three results, glances for 40 seconds, and leaves. An agent reads 20 pages per task, cross-checks dates and numbers, and drops the ones that don't verify.
• A human gets persuaded by the headline. An agent gets persuaded by extractable facts: the date, the source, the concrete figure in the first paragraph.
• A human leaves behavioral signals: time on page, scroll, bounce. An agent leaves none of that, only extraction patterns: how cleanly it can read your HTML, whether schema is there, whether the answer sits before the first 100 tokens.
• A human searches again when convenient. An agent can run the same task every morning at 6am. That turns freshness into a real competitive edge: the page updated on Tuesday beats the page last touched in March.

The agent-readiness check

In the audits we run at Mintec, the problem is almost never "Google can't find me." It's "an agent read me and didn't cite me." Four checks to see whether your page passes:

1. Does the answer extract in a single pass? The agent fetches your URL and parses the raw HTML. If the answer sits behind a cookie modal, a JavaScript wall, or a newsletter popup, the agent loses it. The direct answer needs to be in the first paragraph, plain text, no rendering required.

2. Can every important claim be verified? This generation of agents cross-checks sources. If you write "tool X cut costs 30%," the agent looks for the source of that 30%. No date, no named source, no link to the study, and the claim gets dropped, often along with the whole page. This is the AI search trust gap taken to its extreme: verification is no longer done by a curious human, it's done by a process that runs in seconds.

3. Does your page declare what it is? Schema for Article, FAQPage, Product, or Service. Structured data is what separates a page an agent can classify from one it has to guess about. The same logic applies at site level with llms.txt, and we're increasingly seeing architectures that talk to the agent directly through MCP.

4. Is it alive? Agents weight freshness. A page from January is less likely to appear in a summary in August than the same page updated two weeks ago. You don't need to rewrite everything, just refresh data, dates, and examples.

What our own numbers show

In our weekly GEO traffic report we've tracked LLM referrals for months: they're small, dozens of sessions a week, but they convert about four times better than classic organic. What's new is that citations are starting to arrive without an associated session: pages showing up as sources in ChatGPT or Gemini answers without the user ever visiting. GA4 won't show you that. The Search Console AI reports will, with impressions in AI Overviews and AI Mode already showing how many people see your content without clicking.

On the day Spark launched, our LLM referral traffic was 13 sessions, mostly from ChatGPT, and the most-cited page wasn't the most-visited one on the site. That gap between visits and citations is the metric to watch now. If your strategy still only measures clicks, you're measuring half the funnel.

The LatAm angle: Spark is already in your market

The exclusion list is the most important part of this announcement for our region. When Google leaves the EU and the UK out of a launch, Latin America normally waits months. Not this time: Google AI Pro and Ultra subscribers in Mexico, Colombia, Argentina, Chile, and the rest of the region already have the 24/7 agent live. Local competitors who aren't thinking about agents yet have a window of a few months, not longer. AI isn't going to replace SEO, but it will replace the sites nobody cites.

What to do this week

1. Run a dry fetch on your homepage. Curl your URL and look at the raw HTML: is the answer there? Is there schema? Does any modal block the content? If your homepage fails, agents read half of it and move on.

2. Write the direct answer for your three most important pages. An opening paragraph that answers the question in about 80 words, with a date, a figure, and a source. That paragraph is what the agent will quote. It's the same principle as our 30-minute GEO audit, aimed at agents.

3. Set a freshness alarm. Pick five pages you want agents to cite and schedule a quarterly review of data and dates. Scheduled agents run every day; March content loses to August content even when it's better.

Spark isn't another search engine. It's the first time Google has put a 24/7 worker between your content and your customer. The good news: agents are predictable. If they verify, they cite. And when they cite, the click, when it finally comes, comes already decided.