LLM Optimisation
Inside Google’s AI Overviews: What “folsrch” Reveals About How Synthetic Answers Are Built
There’s a hidden layer behind Google’s new AI Overviews, and it’s quietly reshaping how organic results are generated and how brands need to think about visibility.
2 July 2025
7 min read
While analysing the structure of AI Overviews, we uncovered a recurring internal marker: folsrch. It’s not public-facing, but it’s triggered every time Google attempts to generate an AI Overview — even when it claims one “couldn’t be generated.”
This isn’t just metadata. It’s the blueprint of how synthetic answers are constructed. Here’s what each component does:
🔹 folsrch-sqf — Capturing Intent
This logs the exact query, down to phrasing and semantic cues. It’s the foundation of the pipeline — everything that follows depends on how this input is parsed.
🔹 folsrch-sources — Mapping the Landscape
Google then assembles a curated set of sources: URLs, titles, snippets. These aren’t just citations; they’re used for semantic filtering and fact-checking before generation.
🔹 folsrch-fragment — Extracting Building Blocks
Next, content is sliced into fragments grouped by topic or intent, tagged and attributed. These become the ingredients for the AI response.
🔹 folsrch-response — Synthesising the Output
This is where the final Overview is built: rewritten, condensed, and structured into a fluent, human-like answer.
🔹 folsrch-ghost — Silent Failures
Even when Google says “no overview available,” this ghost marker shows that the pipeline was triggered, but the result didn’t meet quality thresholds.
Why This Matters
Most conversations about AI Overviews focus on what they say. But folsrch tells us how they’re made and that’s where the real optimisation opportunity lies.
At LMO7, we’re helping brands move beyond keywords to semantic architecture: making sure your content is structured, sourced, and surfaced in a way that’s LLM-ready — because increasingly, that’s what’s powering the results users see.
This isn’t just metadata. It’s the blueprint of how synthetic answers are constructed. Here’s what each component does:
🔹 folsrch-sqf — Capturing Intent
This logs the exact query, down to phrasing and semantic cues. It’s the foundation of the pipeline — everything that follows depends on how this input is parsed.
🔹 folsrch-sources — Mapping the Landscape
Google then assembles a curated set of sources: URLs, titles, snippets. These aren’t just citations; they’re used for semantic filtering and fact-checking before generation.
🔹 folsrch-fragment — Extracting Building Blocks
Next, content is sliced into fragments grouped by topic or intent, tagged and attributed. These become the ingredients for the AI response.
🔹 folsrch-response — Synthesising the Output
This is where the final Overview is built: rewritten, condensed, and structured into a fluent, human-like answer.
🔹 folsrch-ghost — Silent Failures
Even when Google says “no overview available,” this ghost marker shows that the pipeline was triggered, but the result didn’t meet quality thresholds.
Why This Matters
Most conversations about AI Overviews focus on what they say. But folsrch tells us how they’re made and that’s where the real optimisation opportunity lies.
At LMO7, we’re helping brands move beyond keywords to semantic architecture: making sure your content is structured, sourced, and surfaced in a way that’s LLM-ready — because increasingly, that’s what’s powering the results users see.