Google Upgrades AI Mode as Search Automation Expands

Visibility Report #30 | August 18, 2026

Google is rolling Gemini 3.7 Flash into AI Mode, changing the system that selects, synthesizes, and cites information for searchers. Meanwhile, OpenAI says the robots.txt rules governing ChatGPT's fetch bot may not work the way publishers expect, and Google has set a migration timeline for AI Max. Now visibility teams have to watch retrieval, crawler access, paid-search automation, and source eligibility at the same time.

Recommended listen: My Search Session conversation with Gianluca Fiorelli pairs well with this issue. We talk through the move from classic rankings to agentic, semantic, and personalized search, including entities, query fan-out, prompt tracking, and why “zero-exit” may describe the new search experience better than “zero-click.”

Google Brings Gemini 3.7 Flash to AI Mode

Google is rolling Gemini 3.7 Flash into AI Mode in Search, making the model update directly relevant to publishers and brands competing for citations. A new model can change query interpretation, answer composition, and source selection even when the underlying index remains similar. Teams should rerun a stable set of prompts and compare cited domains, URLs, and answer language before treating any visibility movement as a content-performance change.

What remains uncertain is how much citation behavior differs from the prior model across commercial, local, and informational queries. Analytics aren't truth. They're opinions with decimal points, and model migrations make those opinions even less stable.

📎 Search Engine Land​

OpenAI Says Robots.txt May Not Control ChatGPT's Fetch Bot

OpenAI says standard robots.txt directives may not apply to the fetch bot ChatGPT uses when a user requests a page. That distinction matters because blocking training or search crawling may not prevent ChatGPT from retrieving a URL in response to a live request. Publishers need to separate model training, search indexing, and user-initiated fetching when setting access policy.

The practical move is to review server logs and confirm which OpenAI user agents reach the site, what they request, and how the server responds. Documentation and observed behavior may still change, so we would avoid assuming one directive governs every OpenAI access path.

📎 Search Engine Journal​

Google Sets the AI Max Migration Timeline

Google has published its migration timeline for moving Search campaigns toward AI Max, giving advertisers a clearer deadline for reviewing controls and reporting. The change affects how queries, creative assets, landing pages, and targeting signals are handled by automation. Search and SEO teams may need a shared process because broader query matching can expose gaps between paid landing pages, organic content, and the language customers use.

Before migration, export current search-term, conversion, asset, and landing-page data. That snapshot will probably be more useful than trying to reconstruct the old baseline after campaign behavior changes.

📎 Search Engine Land​

Candidate Page Selection Is the Missing GEO Measurement Layer

Lumar examines the stage before an AI system retrieves or cites a page: whether that page becomes a candidate at all. Technical accessibility, relevance, internal structure, and extractable passages can influence eligibility before citation quality is considered. This explains why polishing an answer block may do little when the system never places the URL in its retrieval set.

Most visibility tools measure observed answers and citations, not candidate eligibility. Operators should pair prompt monitoring with crawl diagnostics, internal-link review, indexability checks, and page-level content comparisons.

📎 Lumar​

Your Website Still Trains the Market Without Getting the Click

SparkToro argues that websites still matter even as zero-click answers reduce direct visits. The site remains the controlled source where a company can define products, evidence, terminology, and brand claims for search engines and AI tools to interpret. Traffic alone therefore understates the site's role in influencing answers that occur elsewhere.

We think the useful shift is to measure the website as an information source as well as a destination. Track citations, branded answer accuracy, assisted conversions, and downstream searches alongside sessions.

📎 SparkToro​

Claude Watermarks Need Careful Interpretation

Anthropic has disclosed more about the watermarking used in Claude-generated text, prompting questions about authorship detection and search risk. Reporting from Ahrefs and Search Engine Journal indicates the signal has limitations and can be weakened through editing, so it should not be treated as proof that a person or model wrote an entire document. For marketers, usefulness and intent remain more actionable quality tests than attempting to conceal or detect a production method.

The search implication is measured: a watermark does not establish that content manipulates rankings, and its presence does not tell us whether the material is accurate or helpful. Teams should retain editorial review and source verification regardless of how copy was drafted.

📎 Ahrefs | Search Engine Journal​

Google Adds More AI to Ads and Analytics

Google announced new AI features across its advertising and analytics products, extending automation into campaign creation, analysis, and decision support. The useful question is not whether the tools can produce recommendations, but whether operators can inspect the inputs, constraints, and outcome data behind them. Automated suggestions can improve speed while still obscuring budget drift or changes in query coverage.

Before adopting a recommendation, define the metric, acceptable variance, excluded traffic, and rollback condition. That gives teams a governance layer when the platform's optimization objective differs from the business objective.

📎 Google​

Profound Adds Citation Decay Tracking

Profound introduced Citation Decay to show when a brand or page loses citations over time. This helps distinguish a temporary appearance from a durable source relationship, which ordinary share-of-voice snapshots can miss. It may also help teams identify whether losses follow content aging, competitor updates, or changes to an answer engine.

The missing piece is causal certainty: a declining citation does not by itself explain why the model changed sources. Use decay alerts as investigation triggers, then compare the cited page, competing sources, freshness, and prompt output.

📎 Profound​

Repeated Buying Prompts Produce Different Recommendations

Foundation ran the same buying prompt 11 times to examine query fan-out and recommendation variability. The exercise shows why a single prompt run is a weak basis for measuring brand visibility: systems may research different subtopics and return different products on repeated attempts. A mention is also not the same as a recommendation, particularly when the model includes a brand only as an alternative.

In our experience, prompt tracking works better when it uses repeated runs, separates mentions from recommendations, and records supporting citations. They're a windsock, not a GPS.

📎 Foundation Inc​

From the Tool Blogs

Peec AI:AI Shopping Analytics adds monitoring for product discovery inside answer engines. Peec also tested whether company names influence AI judgments, a useful warning that model perception can be affected by signals unrelated to actual product quality.

Semrush: Its free-chatbot keyword research test compares five tools and stresses validating generated ideas against observed search data. That caveat is the value: chatbots can broaden a topic set, but they do not supply dependable demand estimates.

Ahrefs:37 practitioner-submitted AI marketing workflows is useful when paired with human review and measurement, but the watermark analysis belongs in the main section above.

Agency and Practitioner Insights

Search Influence:Why "AI Keeps Changing" Is the Wrong Reason to Skip AI Search Optimization argues that changing interfaces do not remove the need for clear, accessible, authoritative source material.

Foundation Inc:Perplexity cites ClickUp 6,474 times. Notion gets 741... Why? compares citation outcomes for two well-known software brands. The reported gap is useful as a research lead, though teams should inspect prompt selection, time range, and page mix before generalizing it.

Search Influence:Content Chunking Strategy for Higher Ed AI Search connects page structure with retrieval and answer extraction. The advice is especially relevant to institutions with long program pages serving several audiences and intents.

More From This Week

Search and AI Visibility

• ​ChatGPT's Index, GA Benchmarks, Google Refiles Complaint (Search Engine Journal) - A roundup of search developments, including questions about ChatGPT indexing and measurement.
• ​Anthropic Reveals What the Watermark Is and How It Can Be Defeated (Search Engine Journal) - Details on Claude's watermark and the limits of treating it as reliable authorship evidence.
• ​The 4-Step Roadmap to AI Agents for Google Ads (Search Engine Land) - A staged approach to adding agent-driven campaign work while maintaining human controls.
• ​Machine Traffic Could Hit 1,000x Human Traffic in Five Years (Search Engine Journal) - Cloudflare's warning raises crawler governance, security, and infrastructure questions for publishers.

From the Tool Blogs

• ​Three Workflows to Try With the Profound MCP (Profound) - Examples for connecting visibility data with agent and reporting workflows.
• ​The Profound Connector Is Live in Claude - Part 2 (Profound) - Extends access to visibility data inside Claude-supported workflows.

Agency Insights

• ​Traditional SEO Platforms With AI Tracking Features (Search Influence) - Compares how established SEO platforms are adding AI visibility monitoring.
• ​From Cold to Gold: How to Measure Lead Quality (Search Influence) - A measurement framework with limited direct GEO value, but useful for judging whether AI visibility produces qualified demand.

So What Do You Do About It?

Build a fixed benchmark set before the model and campaign migrations obscure the baseline. Run your highest-value prompts several times in Google AI Mode and ChatGPT, record every cited URL and recommendation, export current AI Max search-term and landing-page data, and check server logs for OpenAI user agents. Repeat the same collection after each rollout so you can separate model variability from changes you caused.

The Visibility Report | Will Scott​
This newsletter is produced collaboratively by Will Scott and Bob, an AI agent. Human oversight, AI efficiency.
​UnsubscribeGoogle Upgrades AI Mode as Search Automation Expands

Visibility Report #30 | August 18, 2026

Google is rolling Gemini 3.7 Flash into AI Mode, changing the system that selects, synthesizes, and cites information for searchers. Meanwhile, OpenAI says the robots.txt rules governing ChatGPT's fetch bot may not work the way publishers expect, and Google has set a migration timeline for AI Max. Now visibility teams have to watch retrieval, crawler access, paid-search automation, and source eligibility at the same time.

Recommended listen: My Search Session conversation with Gianluca Fiorelli pairs well with this issue. We talk through the move from classic rankings to agentic, semantic, and personalized search, including entities, query fan-out, prompt tracking, and why “zero-exit” may describe the new search experience better than “zero-click.”

Google Brings Gemini 3.7 Flash to AI Mode

Google is rolling Gemini 3.7 Flash into AI Mode in Search, making the model update directly relevant to publishers and brands competing for citations. A new model can change query interpretation, answer composition, and source selection even when the underlying index remains similar. Teams should rerun a stable set of prompts and compare cited domains, URLs, and answer language before treating any visibility movement as a content-performance change.

What remains uncertain is how much citation behavior differs from the prior model across commercial, local, and informational queries. Analytics aren't truth. They're opinions with decimal points, and model migrations make those opinions even less stable.

📎 Search Engine Land​

OpenAI Says Robots.txt May Not Control ChatGPT's Fetch Bot

OpenAI says standard robots.txt directives may not apply to the fetch bot ChatGPT uses when a user requests a page. That distinction matters because blocking training or search crawling may not prevent ChatGPT from retrieving a URL in response to a live request. Publishers need to separate model training, search indexing, and user-initiated fetching when setting access policy.

The practical move is to review server logs and confirm which OpenAI user agents reach the site, what they request, and how the server responds. Documentation and observed behavior may still change, so we would avoid assuming one directive governs every OpenAI access path.

📎 Search Engine Journal​

Google Sets the AI Max Migration Timeline

Google has published its migration timeline for moving Search campaigns toward AI Max, giving advertisers a clearer deadline for reviewing controls and reporting. The change affects how queries, creative assets, landing pages, and targeting signals are handled by automation. Search and SEO teams may need a shared process because broader query matching can expose gaps between paid landing pages, organic content, and the language customers use.

Before migration, export current search-term, conversion, asset, and landing-page data. That snapshot will probably be more useful than trying to reconstruct the old baseline after campaign behavior changes.

📎 Search Engine Land​

Candidate Page Selection Is the Missing GEO Measurement Layer

Lumar examines the stage before an AI system retrieves or cites a page: whether that page becomes a candidate at all. Technical accessibility, relevance, internal structure, and extractable passages can influence eligibility before citation quality is considered. This explains why polishing an answer block may do little when the system never places the URL in its retrieval set.

Most visibility tools measure observed answers and citations, not candidate eligibility. Operators should pair prompt monitoring with crawl diagnostics, internal-link review, indexability checks, and page-level content comparisons.

📎 Lumar​

Your Website Still Trains the Market Without Getting the Click

SparkToro argues that websites still matter even as zero-click answers reduce direct visits. The site remains the controlled source where a company can define products, evidence, terminology, and brand claims for search engines and AI tools to interpret. Traffic alone therefore understates the site's role in influencing answers that occur elsewhere.

We think the useful shift is to measure the website as an information source as well as a destination. Track citations, branded answer accuracy, assisted conversions, and downstream searches alongside sessions.

📎 SparkToro​

Claude Watermarks Need Careful Interpretation

Anthropic has disclosed more about the watermarking used in Claude-generated text, prompting questions about authorship detection and search risk. Reporting from Ahrefs and Search Engine Journal indicates the signal has limitations and can be weakened through editing, so it should not be treated as proof that a person or model wrote an entire document. For marketers, usefulness and intent remain more actionable quality tests than attempting to conceal or detect a production method.

The search implication is measured: a watermark does not establish that content manipulates rankings, and its presence does not tell us whether the material is accurate or helpful. Teams should retain editorial review and source verification regardless of how copy was drafted.

📎 Ahrefs | Search Engine Journal​

Google Adds More AI to Ads and Analytics

Google announced new AI features across its advertising and analytics products, extending automation into campaign creation, analysis, and decision support. The useful question is not whether the tools can produce recommendations, but whether operators can inspect the inputs, constraints, and outcome data behind them. Automated suggestions can improve speed while still obscuring budget drift or changes in query coverage.

Before adopting a recommendation, define the metric, acceptable variance, excluded traffic, and rollback condition. That gives teams a governance layer when the platform's optimization objective differs from the business objective.

📎 Google​

Profound Adds Citation Decay Tracking

Profound introduced Citation Decay to show when a brand or page loses citations over time. This helps distinguish a temporary appearance from a durable source relationship, which ordinary share-of-voice snapshots can miss. It may also help teams identify whether losses follow content aging, competitor updates, or changes to an answer engine.

The missing piece is causal certainty: a declining citation does not by itself explain why the model changed sources. Use decay alerts as investigation triggers, then compare the cited page, competing sources, freshness, and prompt output.

📎 Profound​

Repeated Buying Prompts Produce Different Recommendations

Foundation ran the same buying prompt 11 times to examine query fan-out and recommendation variability. The exercise shows why a single prompt run is a weak basis for measuring brand visibility: systems may research different subtopics and return different products on repeated attempts. A mention is also not the same as a recommendation, particularly when the model includes a brand only as an alternative.

In our experience, prompt tracking works better when it uses repeated runs, separates mentions from recommendations, and records supporting citations. They're a windsock, not a GPS.

📎 Foundation Inc​

From the Tool Blogs

Peec AI:AI Shopping Analytics adds monitoring for product discovery inside answer engines. Peec also tested whether company names influence AI judgments, a useful warning that model perception can be affected by signals unrelated to actual product quality.

Semrush: Its free-chatbot keyword research test compares five tools and stresses validating generated ideas against observed search data. That caveat is the value: chatbots can broaden a topic set, but they do not supply dependable demand estimates.

Ahrefs:37 practitioner-submitted AI marketing workflows is useful when paired with human review and measurement, but the watermark analysis belongs in the main section above.

Agency and Practitioner Insights

Search Influence:Why "AI Keeps Changing" Is the Wrong Reason to Skip AI Search Optimization argues that changing interfaces do not remove the need for clear, accessible, authoritative source material.

Foundation Inc:Perplexity cites ClickUp 6,474 times. Notion gets 741... Why? compares citation outcomes for two well-known software brands. The reported gap is useful as a research lead, though teams should inspect prompt selection, time range, and page mix before generalizing it.

Search Influence:Content Chunking Strategy for Higher Ed AI Search connects page structure with retrieval and answer extraction. The advice is especially relevant to institutions with long program pages serving several audiences and intents.

More From This Week

Search and AI Visibility

• ​ChatGPT's Index, GA Benchmarks, Google Refiles Complaint (Search Engine Journal) - A roundup of search developments, including questions about ChatGPT indexing and measurement.
• ​Anthropic Reveals What the Watermark Is and How It Can Be Defeated (Search Engine Journal) - Details on Claude's watermark and the limits of treating it as reliable authorship evidence.
• ​The 4-Step Roadmap to AI Agents for Google Ads (Search Engine Land) - A staged approach to adding agent-driven campaign work while maintaining human controls.
• ​Machine Traffic Could Hit 1,000x Human Traffic in Five Years (Search Engine Journal) - Cloudflare's warning raises crawler governance, security, and infrastructure questions for publishers.

From the Tool Blogs

• ​Three Workflows to Try With the Profound MCP (Profound) - Examples for connecting visibility data with agent and reporting workflows.
• ​The Profound Connector Is Live in Claude - Part 2 (Profound) - Extends access to visibility data inside Claude-supported workflows.

Agency Insights

• ​Traditional SEO Platforms With AI Tracking Features (Search Influence) - Compares how established SEO platforms are adding AI visibility monitoring.
• ​From Cold to Gold: How to Measure Lead Quality (Search Influence) - A measurement framework with limited direct GEO value, but useful for judging whether AI visibility produces qualified demand.

So What Do You Do About It?

Build a fixed benchmark set before the model and campaign migrations obscure the baseline. Run your highest-value prompts several times in Google AI Mode and ChatGPT, record every cited URL and recommendation, export current AI Max search-term and landing-page data, and check server logs for OpenAI user agents. Repeat the same collection after each rollout so you can separate model variability from changes you caused.

The Visibility Report | Will Scott​
This newsletter is produced collaboratively by Will Scott and Bob, an AI agent. Human oversight, AI efficiency.
​Unsubscribe​