Skip to content
← Back to Blog
·5 min read·Hass Dhia

AI search visibility tools: track citations as the UI shifts

AI search visibilityAI Overviewscitation trackingbuild vs licenseagency tools

Google is testing a search page with fewer doors into AI. Search Engine Roundtable reports that Google is testing removal of AI Mode from the search tab bar. It is also testing removal of the Ask anything follow-up box inside AI Overviews. This follows earlier reports of AI Mode buttons leaving the Google homepage.

So which AI search visibility tools hold up when the interface keeps changing? The ones that track the answer, not the door. They record which source an engine cited, for which prompt, on which engine, on which date. Everything else is a chart.

AI search visibility is measured against a moving interface

AI search visibility has a measurement habit that follows the interface instead of the answer. The click data shows the cost. In a study published August 5, 2026, researchers found that clicks to sources cited in AI Overviews occur in only about 1% of visits to AI Overviews.

That figure comes from one study of one surface. It still points the same way as the rest of this essay. A report that leads with clicks describes the roughly 1% of AI Overview visits that end in a click on a cited source. A report that leads with an AI Mode segment describes a menu item that Google is testing away.

An agency can act on this today. Move the headline line in a client report from clicks to citations. Keep clicks as a second line. Label each citation row with the engine and the read date, so a changed button never rewrites history. For a deeper treatment of the gap between citations and sales, see AI search visibility tools: separate citations from sales.

The record outlasts the dashboard

On November 14, 2026, Simon Willison shipped datasette-llm 0.1a11 and datasette-vega 0.7. Both are Datasette releases dated that day.

A chart over an export is only as useful as the record behind it.

What a package cannot supply is last quarter. A dashboard started today starts empty. A record of who was cited, and when, is built one dated read at a time. That makes the record, not the code, the asset.

The record also needs notes. Search Console's help page states that a logging error caused a decrease in impressions on the Generative AI performance report for August 13 to August 17, 2026. Someone who pulled that window without the note would read a vendor-side error as a client-side loss. A maintained history carries those annotations. A weekend build does not.

Who makes the decision

Digiday describes the contest in ad tech as shifting to who can make the decisions that deliver results, as AI automates the workflows. Apply that to an agency: the decision about what a client changes next is the part worth keeping in house.

Lenny's Newsletter's Community Wisdom 305 leads with a thread on becoming the bottleneck for AI agents. An owner who checks every agent-drafted visibility report is the step all work waits on. A record with dated coverage and provenance that a buyer can test gives that check something to test against.

One engine's history misleads

A record from a single engine invites a wrong conclusion. In a study published September 6, 2026, covering 180,000 AI answers, ChatGPT and Perplexity cited the same domains only about 11% of the time for identical queries in the same window. Writesonic's study published July 22, 2026 ran 161,286 prompts across four platforms and measured citation overlap with Jaccard similarity. The method, Jaccard similarity, measures how much the cited sources overlap.

A strong showing on one engine says little about another. A visibility report needs a row per engine, or it overstates what it knows.

Rank tracking does not fill the gap either. An IMC 2026 study published October 1, 2026 found that nearly 30% of AIO-cited domains do not appear in the co-displayed first-page results. The paper reads this as a source selection mechanism distinct from Google's ranking algorithm.

There is one lever the data does suggest. A Surfer study published October 22, 2026 found that the more fan-out queries a page ranked for, the more likely it was to be cited in the AI Overview, with a Spearman correlation of 0.77. A correlation is not a cause. It is still a reason to build content briefs around the sub-questions an engine asks, not only the head term. The audit post on citations and clicks covers how to check that against a client's own logs.

When to license, when to build

Build when the question belongs to one client's own data.

License when the value is an accumulated record across engines, with dated annotations and a way to test it. Interface churn raises that value. Each Google test, such as the AI Mode removal, resets a report built on entry points, and leaves a citation record untouched. The post on hiring versus licensing for ranking volatility walks through the same trade for agency pitches.

Before signing anything, an agency or software team can ask a vendor for four things:

  1. A sample export with the engine, prompt, cited domain and read date on every row.
  2. A statement of which engines are covered and from which date each recorder was live.
  3. A list of annotated anomalies, including vendor-side ones like the August 2026 Search Console logging error.
  4. A way to rerun a sample of prompts and compare the result to the stored record.

A vendor that cannot answer these is selling a dashboard. A vendor that can is selling the history.

The licensing inquiry is at smarttechinvest.com/licensing. Three things can be licensed: the owner-and-signals data API, the AI visibility report, and the content engine as a service. Terms are quoted on a call.

Want more insights like this?

Follow along for weekly analysis on brand strategy, market dynamics, and the patterns that separate signal from noise.

Browse All Articles →

Or license our tools.

Related Articles