SEO · October 19, 2026

SEO Visibility Needs an Asset-Level Measurement Stack

By Morgan Ross · Senior Technical Lead

Analytics dashboard displayed on a laptop screen

The new multimodal report exposes a measurement problem SEO dashboards hide

Google’s new multimodal Search reporting is not another reason to add a column called “AI visibility” to an existing dashboard. It is evidence that the unit of SEO measurement is changing from the keyword to the asset and retrieval path.

On September 24, 2026, Google introduced a Web: multimodal filter in both the Search results Performance report and the Generative AI performance report. It covers web searches that use an image through Lens, Circle to Search on Android, an uploaded image, or Chrome’s “Search this image” feature (Google Search Central’s announcement). Google also says to export the data for analysis in other tools.

That sounds like a reporting feature. Operationally, it is a boundary: the user can arrive through a visual object, the system can retrieve a page, and the report may show the page without revealing the textual question that motivated the search. At the same time, Google’s Generative AI report measures impressions in AI Overviews and AI Mode, while its dimensions and aggregation rules differ from the ordinary Web report (Generative AI performance report documentation).

The recommendation is therefore specific: build separate ledgers for text search, visual search, generative impressions, and business outcomes, then join them at the canonical URL and asset level. Do not add their totals together and call the result visibility. A blended number erases the very differences your team needs to act on.

Search now has several retrieval paths to the same page

A page is no longer discovered only because a typed query matched its text. A product photo, a screenshot, a diagram, or a camera image can become the first retrieval object. The destination may still be a normal web page, but the path to that destination is different.

Google’s September announcement names four visual entry points, but the important distinction is broader than the product names. A text search supplies explicit lexical evidence: a query string, a language, and often an intent that can be classified. A multimodal search supplies an image and perhaps additional context. The system has to infer objects, text, scene, similarity, or an implied task before it can choose web results.

Search Console now separates these paths in the interface, but the separation comes with an important loss of observability. Google’s Performance report documentation says the query dimension is unavailable for multimodal traffic because these searches mostly use images rather than text (Performance report dimensions and data groupings). You can group by page, country, device, date, and search appearance, but you cannot read a neat list of “queries” and optimize a title against them.

That limitation changes what a useful diagnosis looks like:

Question	Text-based Web report	Web: multimodal report
What initiated discovery?	A typed search term	An image-assisted search
Query dimension	Available, subject to anonymization and row limits	Not available for this search type
Useful primary unit	Query × page	Page × asset context × device
First optimization question	Does the page answer the intent?	Can the visual asset and destination be recognized and trusted?
Common false conclusion	Low CTR means a title problem	Low visual traffic means images are not indexed

The last row matters. A missing multimodal query does not prove that the image was irrelevant, poorly labeled, or invisible. It proves only that the reporting surface does not expose the user’s visual prompt as text. Treating an absent field as a negative signal is how dashboards turn instrumentation limits into content decisions.

AI impressions are a different event from a click

Google’s Generative AI performance report adds another useful but non-equivalent surface. The report covers AI Overviews and AI Mode, and it groups data by pages, countries, dates, and devices. Its core metric is impressions: links to a site shown in a generative AI feature. It is not a click report, a citation audit, or a conversion report.

Google describes a chart aggregated at the property level and tables whose aggregation changes with the selected dimension. With a page dimension, the table is grouped by page; with country, device, or date, it is aggregated by property. Google also warns that the newest values can be preliminary and that standard limitations, including a 1,000-row limit, apply (Google’s AI performance report documentation).

A page can therefore appear in an AI impression row without producing a visit. A visual search can produce a web click without appearing in a conventional text-query report. An organic click can happen after a result was displayed alongside an AI feature, without the report proving that the AI feature caused the click. These are different events with different denominators.

Use a ledger that preserves those event types:

observed_at, canonical_url, asset_id, surface, country, device,
property_impressions, page_impressions, clicks, sessions,
conversions, source_export, preliminary, notes

At minimum, surface should distinguish web_text, web_multimodal, web_image, and generative_ai. The exact export fields will depend on your pipeline, but the conceptual separation must survive ingestion. If you rename all four surfaces to organic, your warehouse has already destroyed the evidence needed for analysis.

There is a second trap: do not compute AI CTR by dividing clicks from the ordinary Search results report by impressions from the Generative AI report. Those numbers describe different report populations. If you want an outcome ratio, label it honestly as a cross-report association and join it only at a grain that makes sense—usually canonical URL and time period—not as a user-level causal metric.

The research points to retrieval, not markup superstition

The case for an asset-level stack is not just a product announcement. Two 2026 research papers help explain why simplistic optimization recipes fail.

The University of Toronto study What Drives AI Citations? analyzed two 90-day collections containing 158,099 answers and 468,037 citations. Within the same website, page properties explained only 2.8% of citation variation for ChatGPT, while the engine’s retrieval explained 32%. The authors report that every page ChatGPT cited had first been retrieved by its own search, and that the engine rewrote questions into subqueries that often did not coincide with Google or Bing results (Chen and Koudas, 2026).

The practical reading is not “technical SEO no longer matters.” It is that a page cannot be evaluated only by the markup features a publisher controls. Retrieval is an intermediate system state. You need to know which page and which passage can enter the candidate set before debating how the final answer cites it.

SAGEO Arena, published in the KDD 2026 proceedings, makes the same argument from an evaluation perspective. Its benchmark contains 170,000 web documents across nine domains and preserves both body text and structural information such as headings, meta descriptions, and schema markup. The authors find that body-text optimization alone can degrade retrieval, while structural information helps preserve early-stage visibility; their proposed stage-aware approach separates retrieval, reranking, and generation (Kim et al., SAGEO Arena).

Together, these findings support a three-stage measurement model:

Eligibility and retrieval: Can the system discover and select the asset or page for the user’s path?
Reranking and extraction: Does the retrieved document contain a clear, relevant passage or product fact?
Answer, click, and outcome: Is the page cited, visited, remembered, and useful to the business?

Structured data belongs in the first two stages as supporting evidence, not as a magic switch. Google’s own May 2026 resource says the fundamentals remain relevant and emphasizes valuable, unique content, technical accessibility, and useful image and video content—not a special “GEO markup” checklist (Google’s generative AI Search resource). The asset-level stack should measure what happened at each stage rather than infer success from the presence of a schema block.

Build the dashboard around assets and cohorts

The first implementation should be boring. Export the new report manually if necessary, create a stable asset inventory, and establish a weekly baseline before trying to automate every surface.

1. Give every important visual a durable identity

A page URL is not enough for visual SEO. A commerce page may contain a hero image, alternate product angles, a sizing diagram, a video thumbnail, and user-generated images. Assign an asset_id to the images and videos your team expects to influence discovery. Store the asset URL, canonical page, dimensions, format, alt text, visible caption, product or entity, publication date, and last changed date.

A minimal inventory looks like this:

CREATE TABLE search_assets (
  asset_id TEXT PRIMARY KEY,
  canonical_url TEXT NOT NULL,
  asset_url TEXT NOT NULL,
  asset_type TEXT NOT NULL,
  entity_id TEXT,
  alt_text TEXT,
  caption TEXT,
  first_seen DATE,
  last_modified DATE
);

This does not mean Google will report every impression against an image file. It gives your analysts a controlled way to inspect the page that received the event and the visual material that could plausibly have initiated or supported it.

2. Export matched windows, not isolated screenshots

For each week, export ordinary Web data, Web: multimodal data, Image data, and Generative AI data with the same property, canonicalization rules, country scope, device scope, and date window. Record whether the data is preliminary. Keep the raw export immutable and create a transformed table for analysis.

The most useful first cohorts are not “top pages” and “bottom pages.” They are mismatches:

pages with strong text impressions but no multimodal impressions;
pages with multimodal impressions but no clicks;
pages with multimodal clicks but weak downstream sessions;
pages with AI impressions but no ordinary Web growth;
pages with visual assets changed during the observation window;
pages where image traffic is concentrated on mobile or one country.

A mismatch tells you which stage to inspect. A page with multimodal impressions and no click may have an attractive visual result but a weak destination promise. A page with no impressions after an asset refresh may have an eligibility, indexing, or demand problem. The dashboard cannot decide which one from totals alone; the cohort gives an engineer a place to investigate.

3. Use matched tests for changes

Change one class of asset at a time: image format, descriptive caption, visible product data, page layout, or internal linking. Preserve a control group where feasible. Compare pre- and post-change windows long enough to absorb ordinary demand variation, and segment by device and country.

Do not claim that a new alt text caused a rise in AI impressions from a before-and-after chart alone. Google’s reports have rollout behavior, preliminary values, aggregation differences, and hidden query limitations. The honest output is “the treated cohort showed a change after the asset update,” followed by checks for seasonality, indexing, page changes, and concurrent launches.

The research also argues against overfitting content to a single prompt. The SAGEO Arena paper reports that optimization gains degrade under query expansion, simplification, rephrasing, and abstraction. Write the page and its supporting assets for a stable information need, not a single synthetic phrase that may disappear when the retrieval system rewrites the request.

What SEO teams should stop reporting

Three metrics are especially dangerous when presented without their surface and denominator.

“Total visibility.” Adding blue-link impressions, image impressions, multimodal impressions, and AI-feature impressions creates a larger number but no coherent user event. Report the surfaces separately and show their overlap only when the join is explicit.

“AI CTR.” Unless clicks and impressions come from the same report definition and population, the ratio is not CTR. An AI impression can be a displayed supporting link with no click, while a Search result click may be counted in a different aggregation path.

“Schema caused citation.” Schema can improve machine-readable structure and feature eligibility, but the cited research shows retrieval and site-level effects explain more than a checklist of page properties. Measure whether a change improved a defined cohort at a defined stage.

Replace them with a compact scorecard:

KPI	Surface	Decision it supports
Text impressions, clicks, CTR	Web text	Query and landing-page work
Multimodal impressions, clicks	Web multimodal	Visual asset and destination work
Image impressions, clicks	Image search	Image discoverability work
AI-feature impressions	Generative AI	Retrieval and supporting-page work
Qualified sessions and conversions	Analytics/CRM	Commercial value
Asset change-to-outcome lag	Joined cohort	Experiment planning

The scorecard is less impressive than one headline number. It is much harder to misuse.

The operating model: observe the path, then improve the asset

Google’s multimodal report matters because it gives publishers a new observation point, not because it reveals a complete explanation of visual search. Google’s documentation explicitly withholds the text-query dimension for multimodal searches. The AI report exposes impressions but not a universal answer-level citation and conversion chain. Academic work shows that retrieval and model knowledge can dominate page-level explanations.

That combination leads to a disciplined operating loop:

Inventory: map canonical pages to the assets, entities, and evidence they contain.
Separate: store text, image, multimodal, and generative-AI exports as different surfaces.
Cohort: find pages whose surfaces disagree rather than sorting by one total.
Inspect: review the visual asset, page extractability, entity clarity, and destination promise.
Experiment: change one asset or page property, preserve a comparison group, and log the date.
Reconcile: connect Search Console observations to qualified sessions and conversions without pretending the join proves causality.

This is a better response to multimodal Search than adding more speculative metadata. SEO is becoming an observability problem: the system has several retrieval paths, and each path exposes different evidence. Teams that preserve those boundaries can still make concrete decisions. Teams that compress them into one visibility score will have a dashboard that grows while their explanations get worse.

The winning measurement stack is therefore not the one with the most metrics. It is the one that can answer, for a specific page and asset: how was this discovered, what evidence was exposed, what did the user do next, and which part of that path can we responsibly change?

References

← Back to blogs