AIO/AIM/WG Inspector is a Chrome extension that compares, for one query, the organic sources and those of the AI Overview, AI Mode and Web Guide. On Web Guide it rebuilds the full retrieval funnel, checked against the numbers Google itself prints. Everything runs in your browser, with no data sent anywhere.
Google's generative surfaces no longer stop at AI Overviews and AI Mode: a third one, Web Guide, is being tested in the US and outright replaces the organic list with results grouped and rewritten by an LLM. Web Guide has one property no other surface offers: it prints how many documents it read and how many it kept. This page walks through the three surfaces, what remains observable from the browser, and the extension we built to measure it.
For the historical depth, see our February 2026 deep dive.
The only Google surface that prints its own retrieval funnel.
Launched on July 24, 2025 as a Search Labs experiment, Web Guide replaces the ten blue links with thematic groups: a synthesis paragraph, then results clustered by intent, each rewritten by an LLM. A year later it remains opt-in and US-only (Labs, then the Web tab for enrolled users).
Google describes a query fan-out: a custom Gemini breaks the query into parallel sub-queries, then groups results by theme (announcement covered by Search Engine Land). At Moz, Dr. Pete Meyers reports that Web Guide results run on FastSearch with the RankEmbed model, the lightweight retrieval described at the DOJ trial. Ahrefs stresses one click-side difference: unlike AI Overviews, which partly suppress clicks, every Web Guide result remains a clickable link.
Access: the Labs experiment (it then fires on some searches), the Web tab, or the nem=10 parameter to force it. A caveat for measurement: forcing nem=10 multiplies the candidate pool by two to four compared with the native mode.
This is the central finding of our captures: Web Guide goes through the same retrieval pipeline as AI Mode. Same asynchronous endpoint /async/folwr, same parameter core, same session and auth tokens (garc, lro, mstk), same session thread identifier. At equal budget (native versus native), both surfaces reach into candidate lists of the same order of magnitude, and a third to half of the URLs AI Mode cites sit in the corpus Web Guide read. AI Overviews shares none of that layer.
Above that shared plumbing, everything diverges: of 18 URLs carrying an index on both surfaces, zero share the same rank. AI Mode keeps word-for-word extracts of pages, Web Guide rewrites each document as a third-person summary. One retrieval engine, two consolidations, two reading strategies.
Real capture, query "what is a sd card", July 2026.
The green "stated by Google" banner quotes the sentence Google prints on its page: "Deeply analyzed 21 results, 12 selected".
Every document Web Guide reads carries an index number, and the citations in the generated text point at those numbers. That is what makes the funnel measurable document by document, where it had to be inferred on the other surfaces:
The "Top domains" table then aggregates per domain: how many URLs in the corpus, how many shown, how many cited, and how many citation call sites in the text (the Preamble and Expanded columns, depending on whether the citation sits in the synthesis paragraph or in the expanded analysis). The YouTube row is the interesting case: two URLs read, no card displayed, six citations anyway.
Same engine, two ways of answering, and two sets of sources.
The two historical generative surfaces are still captured by the extension, each with its own tab. Both rely on Google's generation, but they don't sit in the same place and don't follow the same citation logic.
The generative block that appears at the top of a normal results page, above (or in place of) the blue links. The user asks for nothing special: they type their query, and Google decides to surface a written answer with a few cited sources on the side.
Context: the usual /search page, organic results still present below.
A conversational mode of its own, reached through the dedicated tab (parameter udm=50). Google builds a longer answer there, often in several parts, with its own citations and a sources column. Follow-up questions can be chained.
Note: these pages have no #rso block, so there are no classic organic results to compare on the page itself.
An update on our February 2026 study.
In February, our first version of the plugin reconstructed a three-layer pipeline for AIO, with the full list of grounding URLs. That is no longer possible. Here is what disappeared, what still holds, and where the information moved.
The Grounding → Pool → Displayed chain (candidates kept, dropped or injected) is no longer observable in the data sent to the browser: the February version's "AIO pipeline view" relied on fields Google no longer sends. In our July captures, the remaining MAGI URLs match the displayed citations one for one, with no hidden domain. The only injection signal still readable is the video (often YouTube) added to the answer.
The capture core still works: the TgQPHd payload markers are now common to all three surfaces (several hundred per capture), and the data-subtree structure that delimits the blocks has not moved. Citations, entities, fragments and sources still extract cleanly. CSS class names keep changing; data-* attributes survive.
The large "read but not shown" set has not vanished from Google, it changed surface. Web Guide reads 141 to 166 documents on our forced queries and displays only 69 to 80, a 50 to 60% invisibility rate that Google states itself. What February inferred from scripts, Web Guide prints in plain words: the pipeline view is no longer relevant for AIO, and it became relevant for Web Guide.
Framing consequence: on AIO and AIM, the tool no longer claims to reconstruct Google's internal selection and focuses on who is cited, where, with which signals. On Web Guide, the full funnel is measurable again, and it is checked against the numbers Google prints.
Organic, AIO, AIM and Web Guide on the same query.
For one query, which domains are visible in organic, in AIO, in AIM, in Web Guide, and which are shared? With four sets, the Venn diagram goes from 7 to 15 non-empty regions. The full overlap remains a minority.
Same query as above: organic (16), AIO (12), AIM (11), Web Guide (18). Each chip filters the table.
On this query, only 5 domains sit in all four sets at once. Web Guide is the largest contributor (18 domains) and keeps 7 to itself: sources no other surface shows. The table below breaks down each domain, one column per surface:
Every Venn region is clickable and filters the table: "WG · 7" isolates, for instance, the domains only Web Guide surfaces. It is the most direct view of what each surface adds to or removes from the classic ranking.
udm=50) and Web Guide carry no full organic block: the comparison uses the AIO capture of the same query as the organic baseline. And ranks never compare across surfaces: the Web Guide corpus index starts at 0 and behaves like a retrieval rank, the AIO/AIM position is an emission counter that can start at 84. Of 18 URLs indexed on both sides, none carries the same number. Beyond simply "who is cited".
Every document Web Guide reads gets a model-written summary, and each card also carries the classic title and meta description of the same URL. The before/after comes for free: you see exactly what the LLM rewrite keeps from a page, and it is not a truncation, it is comparative data extracted from the content.
Documents that feed the generated text without getting a card or a visible link: 3 to 32 per query in our captures. The anchors of those citations also form a list of entities the model itself chose to attribute each claim.
The entities mentioned in the answer text. In AI Overview, entity links often carry no identifier; in AI Mode, Knowledge Graph identifiers appear. Useful to see which entities Google ties a topic to.
Dates attached to sources: absolute on the Web Guide corpus ("Jul 17, 2026"), relative on its cards ("3 months ago"), and visible in AI Mode too. They indicate which freshness Google favours on a query.
The last injection type still visible: a video (often YouTube) added to the answer. We flag it separately, because it does not follow the same logic as text citations.
Everything runs locally, in your browser.
Download the archive, then load the extension in developer mode, in three steps.
In Chrome, go to chrome://extensions and turn on Developer mode (top-right toggle).
Unzip aio-aim-inspector.zip, click "Load unpacked" and select the unzipped folder.
Search on Google: the extension captures the page's AI Overview, the AI Mode (udm=50) version of the same query, and Web Guide if your account has access to it.
The candidates / read / displayed / cited funnel, Google's own sentence alongside, and the domain table with preamble and expanded-analysis citations.
Paste a list of queries, one per line. The extension processes them in the background, one pass per surface per query, then produces a session report.
The four-set Venn diagram (organic, AIO, AIM, Web Guide) and a per-domain table: per-surface ranks, corpus index, fragment, entity, video.
Compare two dated captures of the same query to see what changed: added or removed citations, rank changes, answer text.
The whole comparison exports to JSON and CSV, Web Guide grounding included, for further processing in your own tools.
No data is sent to any server, to RESONEO or to a third party. Local storage only; uninstalling erases everything. Minimal permissions: the extension only reads the active Google results page.
How Google Search's AI answers work under the hood: court testimony, patents, reverse engineering and parsed data.
The integration of generative AI into Google Search follows a phased trajectory, each phase expanding the role of the LLM. The latest one no longer just adds a block: it replaces the organic list.
The three surfaces share the same Gemini model family and a grounding logic, but diverge on almost everything else. This table reflects what we observe in parsed data and confirmed public sources.
| AI Overviews | AI Mode | Web Guide | |
|---|---|---|---|
| UX position | Block at the top of the standard SERP | Dedicated conversational tab (udm=50) |
Replaces the organic block (nem=10 to force it) |
| Trigger | Automatic, at Google's discretion | User-initiated | Labs, Web tab, or URL parameter |
| Conversational | No (one-shot) | Yes, multi-turn | No |
| Citations | Per-card uuid identifiers | Per-card uuid identifiers | Integer indices into a numbered corpus |
| Source metadata | Title, snippet, favicon, position | Same, plus KG entities | Title, per-document LLM summary, date |
| Volume stated by Google | None | None | « Deeply analyzed N results, M selected. » |
| Organic results | Untouched below | None | Gone, 2 or 3 survive in the shell |
On the transport side the boundary sits elsewhere: AI Mode and Web Guide share the same endpoint and session tokens (see section 1), while AI Overviews goes through a separate channel. Same plumbing, different faucets.
All three surfaces follow a RAG pattern: the model searches first, then generates an answer anchored in the retrieved results. Google Cloud defines grounding as "the ability to connect model output to verifiable sources of information," which "reduces the chances of inventing content." Patent US11769017B1 ("Generative Summaries for Search Results", 2023) describes this pipeline: receive the query, select Search Result Documents, extract snippets, synthesize with an LLM. A notable detail: the system has several candidate generative models and picks one dynamically based on query classification.
Dan Petrovic's analysis of this leak showed that Gemini maintains an ordered cache of retrieval results: markers like [6.2] identify the 6th executed sub-query and its 2nd result, a granular index tying each claim to its source. What Web Guide now makes visible client-side points the same way: a numbered corpus, citations pointing at the numbers, and a funnel where every stage can be counted.
In February we documented three layers for AIO: visible sidebar, in-text citations, and grounding URLs hidden in scripts. That third layer is no longer sent to the browser (section 3). The official eligibility rule has not changed: "to be eligible to be shown as a supporting link in AI Overviews or AI Mode, a page must be indexed and eligible to be shown in Google Search with a snippet."
During the DOJ v. Google antitrust remedies trial (April-May 2025), an internal retrieval system called FastSearch was described under sworn testimony. Paragraph 44 of Judge Mehta's Memorandum Opinion states:
RankEmbed is a deep-learning system trained on click-and-query data and human quality rater scores, with "strong natural-language understanding." New in 2026: at Moz, Dr. Pete Meyers reports that Web Guide's results run precisely on FastSearch with RankEmbed. That is consistent with our observations, since Web Guide shares its whole transport with AI Mode, but the detail comes from a secondary source and should be weighed as such.
What the trial confirms, whatever the exact pipeline: the grounding retrieval layer operates independently from traditional web ranking. Our data shows it query after query, many AI sources do not appear in organic, and vice versa.
Query fan-out is the decomposition of a query into parallel sub-queries to broaden retrieval coverage. Liz Reid defined it at Google I/O 2025: "Search recognizes when a question needs advanced reasoning. It calls on our custom version of Gemini to break the question into different subtopics, and it issues a multitude of queries simultaneously." Google describes the same mechanism for Web Guide in its July 2025 announcement.
q= parameter per capture (the user's own), no sub-query field in any payload, and the string "fanout" found in the source code is the name of a CSS animation. The sub-query counts that circulate are not measurable from the browser: treat them as Google's own account, not as extractable data.The fan-out's indirect footprint remains visible in pool width: 15 to 334 candidates depending on query and mode on Web Guide, against 5 to 15 citations on a classic AI Overview.
AI answers are enriched with entity data from several Google knowledge systems:
| Type | Source | Role |
|---|---|---|
| Knowledge Graph | Google KG | Entity disambiguation, factual anchoring (people, places, concepts) |
| Shopping Graph | Google Shopping | Product entities with price, rating and review counts (decodable from AI Mode payloads) |
| Maps / Places | Google Maps | Local entities, geographic anchoring |
| In-text entity links | Google Search | Clickable terms in AIO text linking to internal searches, no MID exposed |
Enrichment varies by surface: AIO sticks to entity links with no identifier, AI Mode exposes real Knowledge Graph MIDs (kgmid=) and also draws from the Shopping Graph and Maps. A constant observation in our captures: not every entity mentioned in the text gets a KG link. The model knows many of them from the grounding text, not from the graph.
Every Gemini version deployed in Search is described by Google as a "custom" variant, fine-tuned for search tasks.
| Date | Product | Model | Key change |
|---|---|---|---|
| May 2023 | SGE | PaLM 2 + MUM | Launch |
| Late 2023 | SGE | Gemini 1.0 custom | 40% latency reduction |
| May 2024 | AI Overviews | Gemini ~1.5 custom | Public US deployment |
| March 2025 | AIO + AI Mode | Gemini 2.0 custom | AI Mode launch |
| May 2025 | AIO + AI Mode | Gemini 2.5 custom | Google I/O 2025 |
| July 2025 | Web Guide | Gemini custom | Thematic clusters in Labs |
| Nov 2025 | AIO + AI Mode | Gemini 3 Pro | Intelligent routing |
| Dec 2025 | AI Mode (global) | Gemini 3 Flash | Default for AI Mode |
| Jan 2026 | AIO (global) | Gemini 3 | Default for AI Overviews |
Since November 2025, Google routes queries between variants: simple questions go to Gemini 3 Flash, harder ones to Gemini 3 Pro. From a parsing standpoint there is no way to detect which model served a given response: no model name appears in network requests, headers or data. The model identifier is entirely server-side.
Official documentation describes overlapping safety layers: quality raters and red-teaming at training, YMYL filters before generation, SpamBrain and SafeSearch at retrieval, grounding verification after generation. AI Mode adds agentic reinforcement learning (with DeepMind) that rewards statements backed by retrieved sources rather than hallucinated, plus runtime detection of vulnerable situations.
The September 2025 Search Quality Rater Guidelines include AI Overviews evaluation examples for the first time: accuracy, attribution of cited sources, usefulness of the summary.
Patent US12013887B2 (granted June 2024) describes an information gain score: a value between 0 and 1 measuring what a document adds relative to what the user already consulted in their session. If active in the AI systems, cited sources should diversify across successive queries on one topic, a logic particularly aligned with AI Mode's multi-turn design and Web Guide's "different angles" clusters.
Whether it runs in production is unconfirmed. But the patent's timing (mid-2024, amid the proliferation of generated content) and its focus on semantic redundancy make it a plausible candidate for filtering result homogenization.
