How AI Mode Notebooks Change GEO Measurement and Reporting

Do private notebooks change what an AI visibility report measures?

Private notebooks change what an AI visibility report measures because the answer may use files that other people cannot access. Here, AI means artificial intelligence: a system that can produce answers from instructions and available material. Google announced notebooks in AI Mode on 19 August 2026. These notebooks can hold public web articles as well as private files, so a notebook result is not always a private-source result.

A buyer should ask where the evidence came from before accepting a score. GEO, or generative engine optimisation, aims to make a brand easier to find and cite in AI answers. If the tester added a brand report first, the answer may show that the file was useful. It does not, by itself, show that a new buyer could find the same brand through the open web.

This distinction matters when choosing a GEO agency, a monitoring tool or a digital PR service. A public citation, an answer shaped by saved preferences and an answer based on an uploaded file are different observations. Each can help a team make a decision. Combining them in one rate can hide which task the product actually performed.

The Google announcement about notebooks in AI Mode describes how selected sources and instructions can move between connected services. The ChatGPT Search guide explains that saved memories, when enabled, can help rewrite search queries. It also describes the use of approximate location. These features are a reason to record the test setup, not evidence that every answer uses every available source.

What are the three source layers in a reproducible audit?

A useful audit separates open-web tests, tests with personal context and tests with selected files before it counts results. These are test conditions, not grades for answer quality. Start with a public-source test in a fresh session, without uploaded files. Treat it as a declared control, not a perfect copy of every new user's experience.

• Open-web test: Use live public pages, without the tester's files or past chats.
• Saved-context test: Note whether the setup includes saved preferences, earlier chats, language or location.
• Private-source test: Use files that require access rights. Check each notebook source, since some may be public pages.

A screenshot is only part of the record. Save the exact question, follow-up, date, market, language, product and sign-in state. Also note whether memory was on and whether the tester selected any files. Give each notebook a version and list which sources have a public URL.

A selected notebook can mix public articles with private notes. Mark each cited source by its actual access level; do not call the whole answer private just because it came from a notebook. If the answer combines layers, mark it as mixed and keep that label visible. If the source is unclear, record that uncertainty instead of guessing.

Do not copy private documents or saved memories into a routine client report. Use source codes that the authorised team can match to a controlled record. Keep only the detail needed to check the result. This audit method is not a substitute for the organisation's own privacy and access rules.

Google's guidance for AI features and websites retains ordinary Search foundations. These include access for crawlers, useful text and structured data that matches the visible page. A crawler is software that visits web pages; structured data gives facts a machine-readable form. Finding a claim in an uploaded file does not show that its public page passes those checks.

Why does one blended AI visibility score mislead buyers?

A blended score can mislead buyers because it hides which tests required prior access to the brand's material. The denominator, meaning the total tests used to calculate a rate, needs a clear definition. A correct sum can still answer the wrong business question. Report the source conditions beside each rate, not in a footnote.

Consider a made-up example, not a campaign result. A brand appears in 6 of 20 open-web tests, 7 of 10 tests with personal context and all 5 selected-file tests. The last group uses a notebook already containing the brand's reports. The combined rate is 18 of 35, or roughly 51 per cent, while the open-web rate is 30 per cent.

Neither rate is false, but they measure different things. A buyer seeking public discovery needs to see the 30 per cent result first. A team testing a staff knowledge tool may care more about whether the selected files were used correctly. Naming the task prevents a useful private answer from becoming a claim about public reach.

A clearer scorecard uses a separate row for each test condition. In the open-web row, record whether the brand appears, whether the citation opens and whether the linked page supports the claim. In the personal-context row, describe the active context and compare the answer with the control. In the selected-file row, check whether the answer reflects the files and respects their access limits.

Keep mixed or unknown-source answers visible in a separate group. State how they affect the total, so another reader can check the arithmetic. A mention does not prove trust, traffic or a sale. Those outcomes need their own evidence.

Two FL PR methods provide practical starting points for this work. The AI search prompt test-set guide uses stable questions, stated conditions and repeated source checks. The AI search source-kit guide brings expert records, public evidence and update history together. These are the agency's published methods, not proof that any specific notebook feature increases public visibility.

How should teams run a controlled source-layer test?

Run the same decision questions under clear control, personal-context and selected-file conditions, then compare like with like. Do not expect a system to repeat every word on each run. Instead, keep the setup stable enough to see what changed. When a condition changes, record it before comparing the next result.

For an initial audit, we suggest 20–40 questions about definitions, choices, implementation, risk and buying decisions. This range is a planning suggestion, not a platform rule or a statistical promise. A prompt is the question or instruction given to the system. Avoid adding the brand name just to ensure that the answer mentions it.

Tag each question with its market, language and expected type of evidence. Agree any follow-up question before the test, so one supplier does not get an easier route than another. FL PR's shared PR and SEO planning guide offers broader context for joining content and media work. The three-run schedule below is also our audit proposal, not a product requirement.

• Control run: Use a fresh session where possible. Turn memory off, add no private files and state the market and language.
• Saved-context run: State the type of memory, preference or location in use. Do not export the private data itself.
• Notebook run: Keep the file list and instructions fixed for the test. Label public links and private files.
• Repeat checks: Run each setup at three separate times. Save all results, including those that do not mention the brand.
• Source checks: Open every cited link. Check the claim and label the source as brand-owned, editorial, distributed, paid or private.
• Next action: Name the person who will fix access, update a page, check an expert claim or review a private file.

Two screenshots from the same account are not a complete experiment. The conversation, location, memory or selected files may have changed between them. If the team cannot establish the setup, keep the answer as a research note. Do not quietly add it to a rate presented as a controlled comparison.

What can digital PR influence when private sources are involved?

Digital PR can improve public evidence, but it cannot control a user's private notebook or promise a citation. The work includes clear source pages, relevant expert input and accurate links to published coverage. First check whether those records can be found and understood on their own. Then test whether the answer uses them correctly.

FL PR's earned-media and GEO guide distinguishes editorial coverage from brand-owned and paid content. Apply that distinction to the audit record. A brand report remains a brand source even when a model cites it. Adding that report to a notebook does not turn it into independent journalism.

Evidence boundary: Google's notebook announcement describes selected sources, while the ChatGPT Search guide describes memory and location context. These documents support recording the source and session conditions. Separating the results into audit rows is our proposed reporting method, not a measurement standard issued by either company.

A useful PR record could contain an original article URL, the named expert and the claim that the article supports. Add the market, language and date of the source check. FL PR's guide to public brand authority in the AI era provides related reading. If a notebook later uses the article, record that use separately from the original media placement.

What should a GEO measurement contract require?

A GEO measurement contract should define the source groups, raw records, separate totals and decision rules. It should also explain how the team will protect private material. Ask the supplier to show one completed test record before agreeing to the reporting format. The buyer should be able to trace a claimed change back to a question, setup and source.

• Terms: Define a brand mention, a working citation, a public source and an answer based on private files.
• Raw record: Include the question, date, market, language, setup, answer summary, brand role and source link.
• Separate totals: Show the number of tests, mentions and checked citations for each setup.
• Privacy: State who may see source records and how long they are kept. Keep personal data out of routine reports.
• Change log: Start a new baseline when the question set or key test conditions change.
• Decision rules: Say when a finding calls for a page fix, new evidence, media work, a test correction or no action.

Watch for a notebook filled with brand files being presented as proof of organic discovery. Other warning signs include hidden memory settings, mixed countries in one rate and source links that no one has opened. Ask for corrections to the test record first. More polished charts cannot resolve an unknown source condition.

Read the open-web control first when the aim is discovery by a new buyer. Use personal-context and selected-file tests for separate questions, such as staff access to product information. A weak result may call for a better public source, a file update or a cleaner test. It does not always call for more content.

Frequently Asked Questions

Review the source and test setup before treating any answer as evidence of public GEO performance.