I see Google’s unveiling of Gemini Intelligence at the May 12 Android Show as a significant step toward an agent-powered future. Announced alongside a new laptop called the Googlebook, Gemini Intelligence is designed as an underlying layer that works across the Android operating system on laptops, phones, watches, and glasses.
The Googlebook makes that vision tangible to me. Built from the ground up around an AI agent, it can understand what is on the screen and act on it. I could point to a date in an email and have the agent schedule a meeting, or select furniture in an app and see how those pieces might look in my living room.
I believe this ability to complete tasks without requiring someone to open a webpage will fundamentally change how people search, discover information, and conduct commerce. Here is how I expect that shift to affect the search industry.
What the shift to an agentic operating system means
Until now, I have viewed search as a familiar sequence: someone has a question or intent, enters it into a search engine, receives a list of links, and chooses one. Earning a prominent position on that list was the prize, and much of the SEO industry was built around winning that click.
Gemini Intelligence starts from a very different assumption. Search intent still exists, but an AI agent can handle the steps between the request and the outcome. It can read pages, complete forms, and increasingly finish the entire task. Instead of visiting a website myself, I may have an agent visit and use it on my behalf.
When I look for an early example, Chrome Auto Browse stands out. Launched in January and built on Gemini 3, it can manage multistep tasks such as researching flights, filling out forms, scheduling appointments, and managing subscriptions. It then pauses for approval before making a purchase.
A 2025 preprint supports this view. Researchers evaluated the declared-tools approach across online shopping, authentication, and content management. They found that giving an agent pre-structured interaction data reduced processing requirements by 67.6% and lowered costs by 34% to 63% compared with parsing a complete HTML document. Task success declined only slightly, from 98.8% with the traditional method to 97.9%.
The architecture behind Gemini Intelligence
To me, the architecture is as important as the interface. AI agents naturally favor websites they can interact with cleanly and efficiently, and Gemini Intelligence can only deliver on its promise if those agents can perform tasks reliably.
I see two protocols as central to making that possible. WebMCP turns a website’s actions into callable tools, while the Universal Commerce Protocol (UCP) allows an agent to complete a sale. Together, they enable an agent to finish a task without requiring a person to load and navigate the underlying webpage.
As AI reshapes search, every glowing path to discovery carries commercial value—turning SEO investment into a conversation about pipeline, risk, and customer acquisition costs.
WebMCP
I think of WebMCP as a labeled menu for AI agents. The API allows a website to declare functions as structured tools an agent can call, including searching inventory, beginning checkout, or submitting a support request.
Google co-developed WebMCP with Microsoft. An origin trial is live in Chrome 149, Firefox has committed to the third quarter of 2026, and Safari is expected to follow in the fourth quarter.
Universal Commerce Protocol (UCP)
I see UCP as the transactional counterpart to WebMCP. It gives AI agents a shared language for discovering products, building a cart, completing checkout, and managing orders without requiring someone to visit the merchant’s website.
Google also offers a consumer-facing layer called Universal Cart. It can collect items as I move across Search, Gemini, YouTube, and Gmail, creating a more connected shopping experience across Google’s products.
The range of companies behind UCP shows me how seriously the industry is taking this shift. Google, Shopify, Walmart, Target, Etsy, Wayfair, PayPal, and Stripe co-developed the protocol, which launched in January.
How I would prepare for agentic AI
My main takeaway is that websites are rapidly evolving from destinations into backends—from places people actively visit into systems agents quietly use. As the operating system becomes a search and action layer, I no longer think ranking is the only question that matters. I also need to ask whether an agent can actually use the site.
To prepare, I would begin by auditing the site’s most valuable actions, whether that means submitting a lead form, completing a booking flow, or reaching checkout. I would determine whether an agent could complete each action reliably and check the site’s Lighthouse Agentic Browsing score much as I would review Core Web Vitals. The goal is to understand whether an agent can use the site, not merely read it.
If I ran an ecommerce business, I would confirm whether the checkout process is accessible through UCP or ACP. I would also continue investing in retrieval and visibility because an agent still needs to find and trust the business before it can act on anyone’s behalf.
A year ago, I watched the industry place its bets on which AI platform would own discovery. Perplexity looked like the search-native challenger. Copilot looked like the enterprise Trojan horse. In the data I’m seeing now, neither bet has really paid off.
Previsible (disclosure: I’m its CPO and co-founder) just published its third AI Traffic Study, based on 6.77 million LLM-driven sessions. What stands out to me is the level of consolidation. Monthly LLM sessions grew 9.9x, reaching 644,478 in May 2026, and 92.4% of that traffic came from one platform.
The plateau was a pause
In mid-2025, it looked like AI traffic might be topping out in some sectors. I don’t think that was the real story.
Sessions climbed from 65,249 in November 2024 to 396,278 by August 2025. Then they dropped sharply in November 2025 before reaching new highs of 428,203 in February 2026 and 644,478 in May.
That November dip deserves context.
Sessions fell 50% in a single month, driven almost entirely by ChatGPT referrals dropping from 448,412 to 213,345. Other platforms were mostly steady. To me, that points to a model-related change. We’ve already seen small product shifts create major swings in referral traffic, including last fall, when many sites lost half their ChatGPT traffic because the model began favoring Wikipedia and Reddit. By December, sessions had recovered to 442,609.
The lesson I take from this is simple: one vendor’s product decision can cut your AI traffic in half overnight. I would plan for that volatility instead of treating AI referrals as a stable channel.
Consolidation, not competition
When we last published in December 2025, ChatGPT held about 84% share. Perplexity followed at 8.9%, Gemini at 4.5%, Copilot at 2.1%, and Claude at 0.6%. Six months later, the field had moved even more decisively toward the leader.
Across the full dataset, ChatGPT now commands 92.4% of trackable LLM referral traffic. It grew 12.8x over 19 months, with no clear sign of slowing. It is the only LLM sending meaningful referral volume at scale, which means I would not talk about “AI visibility” without putting ChatGPT first.
There is one important caveat. This study measures standalone LLM referral traffic. AI discovery inside Google’s own results, including AI Overviews, almost certainly drives more AI traffic than all standalone platforms combined. But that operates under a different measurement model, so it is not included here.
The challengers flipped
The surprise is not that ChatGPT is on top. What I find more interesting is the movement beneath it.
Claude
Claude grew 64x, moving from 133 sessions in November 2024 to 8,528 in May 2026. It overtook Perplexity in March 2026 for the first time, and it stayed ahead.
Claude was mostly flat through 2025, then accelerated 4x in two months as its agentic tools and enterprise integrations gained adoption. The enterprise advantage many people expected Copilot to win may be materializing for Claude instead.
If your audience includes technical buyers, developers, or professional services, I would treat Claude visibility as material now. The early positioning window is still open, but it may not stay that way for long.
Gemini
Gemini is the quiet number two in this dataset. It delivered 3.2x growth with very little volatility. Because Gemini is tied into Workspace and Android, I suspect referral numbers undercount its real discovery footprint.
Perplexity & Copilot
Perplexity peaked at 17,507 monthly sessions in March 2025 and has fallen 61% since. Copilot fell even harder, dropping 96% from its August 2025 peak, from 8,651 sessions to 339.
I no longer see either platform as a strong traffic-acquisition growth bet. Both are shifting toward experiences that keep users inside their own environments, including browsers, agents, and modes where they do not need to send traffic out at all.
Where LLMs send users, and why it should change your roadmap
The most actionable finding in the study is not market share. It is where LLMs send people after they decide a site is worth visiting.
ChatGPT sends 28.8% of its traffic to internal search results pages. Across industries, roughly 25% of AI-referred traffic lands on internal search.
Old search marketing tools give way to a faster, connected future, with data streams, AI icons, and a glowing search hub symbolizing SEO innovation and community growth.
My read is that the model trusts the domain but cannot always identify the exact right page. So it sends users to the site’s search box and lets them navigate from there. Because this pattern holds across verticals and time periods, I see it as structural to retrieval-augmented generation rather than a temporary quirk.
That changes the role of internal search. The model already did the hard work of choosing your domain. Now your internal search experience decides whether that high-intent visit converts or bounces.
For most sites, internal search is still treated like a neglected navigation feature. I think it needs to be treated as an acquisition surface.
The vertical-level data tells several different stories. SaaS traffic lands on search pages 34.6% of the time. Publisher traffic lands on news pages 54% of the time, but against 120+ million organic sessions, publisher penetration is only 0.11%. Publishers create the content LLMs cite, yet they capture almost none of the resulting traffic.
Ecommerce traffic tends to land on product pages, often with purchase intent already formed. Education traffic lands directly on course pages 52% of the time, bypassing marketing content. Health traffic lands on About pages 42.1% of the time, suggesting users are evaluating the source before trusting the content. Legal traffic spreads across blog, about, contact, and location pages, which reflects the full evaluation arc.
The platforms have distinct behaviors, too. ChatGPT and Gemini act more like search-pattern models: they show domain trust but page-level uncertainty. Perplexity and Claude behave more like content-selection models, picking specific pages and over-indexing on long-form content.
If your strategy depends on editorial content driving qualified traffic, I would give Perplexity and Claude more attention than their raw share suggests.
What I would do now
First, I would optimize for ChatGPT before anything else and expand to other platforms only when the volume justifies the work. ChatGPT is where the measurable standalone LLM referral traffic is concentrated.
Second, I would monitor Claude closely. It overtook Perplexity in March 2026, and early visibility advantages can compound quickly when a platform is still forming its citation and recommendation patterns.
Third, I would treat product pages as AI entry points. Product pages capture 43% of ecommerce LLM traffic, which makes structured, comparable product data a discoverability requirement rather than a nice-to-have.
Fourth, I would make pricing machine-readable wherever possible. “Contact us for pricing” gives AI systems very little to summarize, compare, or recommend.
Fifth, I would prioritize internal search. It is not just a navigation feature anymore. For AI-referred users, it may be the first real conversion point.
Finally, I would track AI traffic by page type instead of relying only on site-wide averages. Your overall AI traffic number can hide where the real concentration is. A pricing page, for example, might run 3x your site-wide penetration.
The next question I want answered is conversion rate by LLM platform. Which platforms send users who buy, and which send users who bounce?
We built this dataset to answer that. If the last 19 months are any guide, I expect the answers to change faster than most teams are prepared for.
About the data
This analysis includes 166 GA4 properties from November 2024 through May 2026, spanning SaaS, ecommerce, finance, legal, health, insurance, education, publishing, and ticketing. All 166 properties are present throughout the full 19-month window, so I’m looking at behavioral change rather than sample expansion.
I see vibe coding as one of the most accessible ways to create small pieces of software with AI tools like ChatGPT, Cursor, Replit, and Gemini. Instead of writing code line by line, I describe what I want in plain language, receive working code in return, paste it into an environment like Google Colab, run it, and test the result.
In this guide, I’ll explain how I approach vibe coding, where I think it works well, where it breaks down, and which SEO examples can inspire practical projects of your own.
Vibe coding variations
I use “vibe coding” as a broad term, but it helps to separate it from nearby approaches:
Type
Description
Tools
AI-assisted coding
AI helps write, refactor, explain, or debug code. I usually associate this with developers or engineers who already understand the systems they are building.
GitHub Copilot, Cursor, Claude, Google AI Studio
Vibe coding
AI handles most of the work after I provide the idea or prompt.
ChatGPT, Replit, Gemini, Google AI Studio
No-code platforms
Platforms handle what I ask for through visual interfaces, drag-and-drop workflows, or background automation. Many now use AI, but they existed before AI became mainstream.
Notion, Zapier, Wix
For this guide, I’m focusing only on vibe coding.
The barrier to entry is low. In most cases, I only need a ChatGPT account, free or paid, and access to a Google account. Depending on the project, I may also need API access or subscriptions to SEO tools such as Semrush or Screaming Frog.
I also like to set expectations early: by the end of this kind of workflow, I’m usually aiming to run a small program in the cloud. If I want to build a SaaS product or software I plan to sell, I treat AI-assisted coding as the more realistic path because it usually requires more technical knowledge, more testing, and more budget.
Vibe coding use cases
I find vibe coding most useful when I’m working with clear buckets of data and need a helpful outcome, not a perfect one. That might mean finding related internal links, adding pre-selected tags to articles, comparing groups of URLs, or building something playful where the output does not need to be exact.
For example, I built an app that creates a daily drawing for my daughter. I type a short phrase about something she told me, such as “I had carrot cake at daycare.” The app uses examples of drawing styles I like and a few pictures of her, then generates a drawing as the final output.
When I ask for precise changes, the tool often gets worse. I once asked it to remove a mustache, and instead it recolored the image. That is exactly the kind of limitation I expect with this approach.
If my daughter were a client reviewing every detail, I would need someone with Photoshop or similar skills to make exact edits. For this use case, though, the result is good enough, and that is where vibe coding shines.
I would be cautious about building commercial applications solely through vibe coding. Some companies may even need vibe coding cleaners to clean up AI-generated work. But for demos, MVPs, internal tools, and quick experiments, I see vibe coding as a useful shortcut.
How I create SEO tools with vibe coding
When I create an SEO tool with vibe coding, I usually follow three steps:
I write a prompt describing the code I need.
I paste the code into a tool such as Google Colab.
I run the code and check whether the results match what I expected.
Here’s a real prompt example from a tool I built to map related links at scale. After crawling a website with Screaming Frog and extracting vector embeddings through the crawler’s OpenAI integration, I vibe coded a tool to compare topical distance between the vectors for each URL.
This is exactly what I wrote in ChatGPT:
I need a Google Colab code that will use OpenAI to:
Check the vector embeddings existing in column C. Use cosine similarity to match with two suggestions from each locale (locale identified in Column A).
The goal is to find which pages from each locale are the most similar to each other, so we can add hreflang between these pages.
I’ll upload a CSV with these columns and expect a CSV in return with the answers.
After ChatGPT generated the code, I pasted it into Google Colab, which is a free Jupyter Notebook environment for running Python in a browser. I then used “Run all” to test whether the program produced the output I wanted.
That is the clean version of the process. In practice, AI can make the workflow look perfect while still producing code that does not behave the way I need.
I expect issues along the way, and most of them are simple to troubleshoot if I keep the prompt and testing process clear.
First, I always state the platform I’m using. If I want code for Google Colab, I say that directly in the prompt.
Sometimes I still get code that depends on packages that are not installed. When that happens, I paste the error back into ChatGPT and ask it to fix the code or suggest an alternative. I do not need to fully understand the missing package to move forward. I can also ask Gemini inside Google Colab to identify the problem and update the code directly.
I also check outputs carefully because AI can sound confident while inventing data. One time, I forgot to specify that the source data would come from a CSV file, so the tool created fake URLs, traffic, and graphs. “It looks good” is not the same as “it is correct.”
If I connect to an API, especially a paid API from a provider such as Semrush, OpenAI, Google Cloud, or another platform, I need to request my own API key and keep usage costs in mind.
A Semrush subscription screen highlights 2 million Standard API units and a masked API key, underscoring the setup step needed for SEO automation and vibe-coded tools.
If I want an even lower execution barrier than Google Colab, I can use Replit.
With Replit, I can prompt what I want, and the platform can generate the code, design the interface, and let me test everything in one place. That reduces copy-and-paste work and gives me a shareable URL quickly. I still need to review poor outputs and keep iterating until the app behaves properly.
The tradeoff is cost. Google Colab is free unless I use paid API keys, while Replit charges a monthly subscription and usage-based API fees. The more the app runs, the more expensive it can become.
SEO vibe-coded tools that inspire me
Google Colab is the easiest place for me to start, but SEOs are taking vibe coding much further. I’ve seen people create Chrome extensions, Google Sheets automations, and even browser games.
I’m sharing these examples because they show what is possible when useful SEO ideas meet practical AI tooling. If I see a tool and wish it had a different feature, that is often a sign that I could try building a version for myself.
A Replit vibe-coding session turns a simple prompt for toddler-friendly daily jokes into a published web app, illustrating how AI tools can quickly prototype playful ideas.
GBP Reviews Sentiment Analyzer by Celeste Gonzalez
After vibe coding SEO tools in Google Colab, Celeste Gonzalez, Director of SEO Testing at RicketyRoo Inc, pushed the idea further by creating a Chrome extension. “I realized that I don’t need to build something big, just something useful,” she explained.
Her extension, the GBP Reviews Sentiment Analyzer, summarizes sentiment analysis from reviews over the last 30 days and shows review velocity. It also exports the information to CSV and works on Google Maps and Google Business Profile pages.
Instead of relying only on ChatGPT, Celeste used Claude to create stronger prompts and Cursor to turn those prompts into code.
AI tools used: Claude (Sunner 4.5 model) and Cursor
APIs used: Google Business Profile API (free)
A vibe-coded GBP Sentiment Analyzer turns review data into a quick snapshot, showing negative sentiment trends, key topics, and an export option for SEO workflows.
Platform hosting: Chrome Extension
Knowledge Panel Tracker by Gus Pelogia
I became obsessed with the Knowledge Graph in 2022, when I learned how to create and manage my own knowledge panel. Later, I discovered that Google’s Knowledge Graph Search API lets me check the confidence score for any entity.
That led me to build a vibe-coded tracker that checks entity scores daily, or at any frequency I choose, and returns the results in a Google Sheet. I can track multiple entities at once and add new ones whenever I need to.
The Knowledge Panel Tracker runs entirely in Google Sheets, and the Knowledge Graph Search API is free to use. This guide explains how to create and run it in your own Google account, or you can see the spreadsheet here and update the API key under Extensions > App Scripts.
AI models used: ChatGPT 5.1
A spreadsheet-based Knowledge Panel Tracker turns entity searches into structured SEO data, comparing names, entity types, descriptions, and confidence scores at a glance.
APIs used: Google Knowledge Graph API (free)
Platform hosting: Google Sheets
Inbox Hero Game by Vince Nero
I also like the idea of vibe coding a link building asset. That is what Vince Nero from BuzzStream did with the Inbox Hero Game. The game asks players to use the keyboard to accept or reject a pitch within seconds, and it ends if they accept too many bad pitches.
Inbox Hero Game is more complex than running a small script in Google Colab, and it took Vince about 20 hours to build from scratch. “I learned you have to build things in pieces. Design the guy first, then the backgrounds, then one aspect of the game mechanics, etc.,” he said.
The game was built with HTML, CSS, and JavaScript. “I uploaded the files to GitHub to make it work. ChatGPT walked me through everything,” Vince explained.
A retro arcade-style Inbox Hero screen turns PR pitch triage into a fast keyboard game, challenging players to accept or reject emails before time runs out.
He also found that longer prompt threads became less useful over time, “to the point where [he’d] have to restart in a new chat.”
That became one of the hardest parts of the project. Vince would add a feature, such as a score, and ChatGPT would “guarantee” it had found the error, update the file, and still return the same problem.
In the end, Inbox Hero Game shows that it is possible to create a simple game without coding knowledge. It also shows where a developer becomes valuable when the goal shifts from “working prototype” to polished product.
AI models used: ChatGPT
APIs used: None
Platform hosting: Webpage
How I think about vibe coding with intent
I do not expect vibe coding to replace developers, and I do not think it should. What I do see is a practical way for SEOs to prototype ideas, automate repetitive tasks, and explore creative experiments without a heavy technical lift.
The key is realism. I use vibe coding where precision is not mission-critical, I validate outputs carefully, and I stay alert for the moment when a project grows beyond “good enough” and needs stronger technical support.
When I approach vibe coding thoughtfully, it becomes less about shipping perfect software and more about expanding what I can test. For internal tools, proofs of concept, and SEO side projects, the best results come from pairing curiosity with restraint.
I’m seeing Google roll out a new set of Demand Gen updates designed to help advertisers improve creative performance, reach more potential customers across YouTube, and measure campaign results with more clarity.
For me, the bigger story is that Demand Gen is becoming less about manually adapting assets and more about using AI-assisted tools to make creative work harder across Google’s most visual surfaces.
Demand Gen campaigns are built to drive discovery and conversions across Google’s visual placements. With these latest updates, I see Google trying to reduce creative friction while giving advertisers better visibility into what is actually moving performance.
Google says the enhancements arrive as YouTube continues to show value for customer acquisition. The company cited research from Measured showing that 72% of incremental conversions on YouTube come from new customers.
What’s new. I’m watching Demand Gen add expanded video resizing capabilities, giving advertisers the ability to automatically transform creative into more aspect ratios, including vertical-to-square, vertical-to-landscape, and square-to-landscape formats.
That matters because it should make it easier to adapt existing creative for different YouTube placements without having to produce every version manually from scratch.
Why I care. Expanded video resizing can help existing assets fit more YouTube inventory, Gemini can provide AI-powered recommendations before launch, and new web-to-app measurement can give marketers a clearer view of how Demand Gen campaigns influence app installs and return on ad spend.
Gemini joins the creative workflow. Google is also bringing Gemini-powered recommendations directly into the Demand Gen campaign creation process, which makes AI guidance part of the asset selection workflow instead of a separate optimization step.
When advertisers choose image and video assets, Gemini will offer automated suggestions for optimizing creative for YouTube. I see this as a way for marketers to improve asset choices before campaigns go live, rather than waiting for performance data after launch.
Better app measurement. Demand Gen now includes Web to App Acquisition Measurement, allowing advertisers to measure when web campaigns lead users to install an app.
The new reporting gives me a more complete way to evaluate campaign performance because it attributes app installs generated through Demand Gen campaigns. That should help advertisers better understand the full impact of their media spend.
The bottom line. I see Google’s latest Demand Gen updates as a practical combination of AI-powered creative guidance, more flexible video optimization, and broader measurement tools that can help advertisers improve performance while gaining clearer insight into customer acquisition.
I have watched the debate around llms.txt become one of the most polarized conversations in web optimization.
Some people treat llms.txt as essential infrastructure for AI discovery. Others, especially longtime SEO practitioners, see it as speculative theater. Platform tools are starting to flag missing llms.txt files as site issues, yet server logs still show that AI crawlers rarely request them.
Google even appeared to adopt it. Sort of. In December, Google added llms.txt files across many developer and documentation sites.
At first, the signal looked obvious to me: if the company behind the sitemap standard was implementing llms.txt, maybe the file really mattered.
Then Google removed it from its Search developer docs within 24 hours.
Google’s John Mueller said the change came from a sitewide CMS update that many content teams didn’t realize was happening. When asked why the files still exist on other Google properties, Mueller said they aren’t “findable by default because they’re not at the top-level” and “it’s safe to assume they’re there for other purposes,” not discovery.
The llms.txt research
I wanted data, not another debate.
So I tracked llms.txt adoption across 10 sites in finance, B2B SaaS, ecommerce, insurance, and pet care. I looked at the 90 days before implementation and the 90 days after.
I measured AI crawl frequency, traffic from ChatGPT, Claude, Perplexity, and Gemini, and the other changes each site made during the same window.
Here is what I found:
Two of the 10 sites saw AI traffic increases of 12.5% and 25%, but llms.txt was not the cause.
Eight sites saw no measurable change.
One site declined by 19.7%.
The 2 ‘success’ stories weren’t about the file
The Neobank: 25% growth
One digital banking platform implemented llms.txt early in Q3 2025. Ninety days later, its AI traffic was up 25%.
That sounds compelling until I looked at what else happened during the same period.
The company ran a PR campaign around its banking license and earned coverage in major national publications.
It restructured product pages with extractable comparison tables for interest rates, fees, and minimums.
It published 12 new FAQ pages optimized for extraction.
It rebuilt its resource center with new banking information and concepts.
It fixed technical SEO issues, including header structure problems.
When a company earns Bloomberg coverage in the same month it launches optimized content and fixes crawl errors, I cannot isolate llms.txt as the growth driver.
The B2B SaaS platform: 12.5% growth
A workflow automation company saw AI traffic jump 12.5% two weeks after implementing llms.txt.
The timing looked perfect. It would be easy to call the case closed. But the surrounding context told a different story.
Three weeks earlier, the company had published 27 downloadable AI templates covering project management frameworks, financial models, and workflow planners. These were functional tools, not ordinary content marketing assets, and they drove the engagement behind the spike.
Google organic traffic to those templates rose 18% during the same period and kept climbing throughout the 90 days I measured.
Search engines and AI models surfaced the templates because they solved real problems and created an entirely new site section. They did not surface them simply because the URLs appeared in an llms.txt file.
The 8 sites where nothing happened after uploading llms.txt
Eight sites saw no measurable change after adding llms.txt. One of them declined by 19.7%.
The decline came from an insurance site that implemented llms.txt in early September. Based on the data, the drop likely had nothing to do with the file.
The same pattern appeared across all traffic channels. Llms.txt did not prevent the decline, and it did not create any visible advantage.
The other seven sites, which included ecommerce brands in pet supplies, home goods, and fashion, plus B2B SaaS, finance, and pet care sites, used llms.txt to document their best existing content. That content included product pages, case studies, API docs, and buying guides.
Ninety days later, nothing changed. Traffic stayed flat. Crawl frequency was identical. The content was already indexed and discoverable, and the file did not change that.
The pattern was clear: sites that launched new, functional content saw gains. Sites that only documented existing content saw no gains.
Why the disconnect?
No major LLM provider has officially committed to parsing llms.txt. Not OpenAI. Not Anthropic. Not Google. Not Meta.
“None of the AI services have said they’re using llms.txt, and you can tell when you look at your server logs that they don’t even check for it.”
That is the reality I saw in the data. The file exists. The advocacy exists. But platform adoption does not show meaningful use yet.
The token efficiency argument and its limits
The strongest case for llms.txt is efficiency. Markdown can save time and tokens when AI agents parse documentation. It gives agents clean structure instead of forcing them through complex HTML, navigation, ads, and JavaScript.
That matters, but mostly for developer tools and API documentation. If your audience uses AI coding assistants like Cursor or GitHub Copilot to interact with your product, token efficiency can improve integration.
For ecommerce brands selling pet supplies, insurance companies explaining coverage, or B2B SaaS companies targeting nontechnical buyers, token efficiency does not automatically translate into traffic.
llms.txt is a sitemap, not a strategy
The closest comparison I can make is a sitemap.
Sitemaps are useful infrastructure. They help search engines discover and index content more efficiently. But I would not credit traffic growth to simply adding a sitemap. The sitemap documents what exists; the content drives discovery.
Llms.txt works in a similar way. It may help AI models parse a site more efficiently if they choose to use it, but it does not make the content more useful, authoritative, or likely to answer user queries.
In my analysis, the sites that grew did so because they:
Created functional assets such as downloadable templates, comparison tables, and structured data.
Earned external visibility through press and backlinks.
Fixed technical barriers such as crawl and indexing issues.
Published content optimized for extraction, including FAQs and structured comparisons.
Llms.txt documented those efforts. It did not drive them.
What actually works
The two successful sites showed me what actually matters.
Create functional, extractable assets. The SaaS platform built 27 downloadable templates that users could deploy immediately. AI models surfaced them because they solved real problems, not because they appeared in a markdown file.
Structure content for extraction. The neobank rebuilt product pages with comparison tables for interest rates, fees, and account minimums. That is data AI models can pull directly into answers without heavy interpretation.
Fix technical barriers first. The neobank fixed crawl errors that had blocked content for months. If AI models cannot access your content, no amount of documentation will help.
Earn external validation. Coverage from Bloomberg and other major publications drove referral traffic, branded searches, and likely influenced how AI models assessed authority.
Optimize for user intent. Both sites answered specific queries, such as “best project management templates” and “how do [brand] interest rates compare?” Models surface content that maps to what users ask, not content that is merely well documented.
None of this requires llms.txt. All of it can drive results.
Should you implement an llms.txt file?
If you run a developer tool and AI coding assistants are a primary distribution channel, I would implement llms.txt. In that context, token efficiency matters because your audience is already using agents to work with documentation.
For everyone else, I would treat llms.txt like a sitemap: useful infrastructure, not a growth lever.
It is good practice to have. It likely will not hurt. But the hour spent implementing llms.txt is often better spent restructuring product pages with extractable data, publishing functional assets, fixing technical SEO issues, creating FAQ content, or earning press coverage.
Those tactics have shown real ROI in AI discovery. Llms.txt has not, at least not yet.
The lesson I take from this is not that llms.txt is bad. It is that we are reaching for control in a system where the rules are still being written. Llms.txt offers comfort because it is concrete, actionable, and familiar. It looks like the web standards we already understand.
But looking like infrastructure is not the same as functioning like infrastructure.
My focus would stay on what is already working:
Create useful content.
Structure it for extraction.
Make it technically accessible.
Earn external validation.
Platforms and formats will change. The fundamentals will not.
Have you ever found yourself immersed in the SEO world, only to be told by an AI that everything you know is wrong? That’s exactly what happened to me, and not just once, but three times in a single week with Gemini.
It’s not the mistakes that rattled me—it was how credible they sounded. The answers from Gemini were polished and convincing, enough so that most would accept them without question.
When it comes to topics you’re not deeply versed in, how do you even begin to challenge such confident wrongness?
Laughably, I caught two, but the third one hit me where it hurts—my wallet. All this unfolded within a week.
Here’s a closer look at what went down.
In one scenario, Gemini misguidedly walked me through technical SEO for a client. During a site migration task on Shopify, where canonical tags were misbehaving, I turned to Gemini for solutions.
The advice was not just misleading but used terms that would raise red flags with leadership—talk about penalties!
Semantic clarity is crucial here; an internal misstep with jargon can make stakeholders halt essential projects.
Gemini further compounded the issue with incorrect guidance on URL parameters hosting.
The experience echoes another incident where Gemini’s mechanical advice almost led me to make a $3,000 error on my Jeep SRT. The AI’s confident proclamation of a rear differential issue had me nearly misappropriating my resources.
After sharing more data, Gemini pivoted, claiming it had leapt to conclusions without sufficient evidence.
In yet another amusing episode, my Madden game finance strategy, courtesy of Gemini, resulted in a fictional $20 million oversight. Although the stakes were virtual, it was a stark reminder of why critical thinking remains indispensable.
These anecdotes underline that it’s not AI replacing experts but rather pushing out those who stop questioning.
The real skill remains in smelling the bull and asking deeper, more insightful questions.
Attending Google I/O 2026 for the first time felt like stepping into a realm of boundless energy and optimism, almost as thrilling as witnessing a crowning ceremony.
The initiatives launched last year have transformed into robust pillars of growth. Ask Maps, for instance, has become the blueprint for introducing Ask YouTube. Gemini 3.5 Flash fuels Antigravity, akin to Claude Code but under Google’s banner, and Googlers are already harnessing it to construct the exciting features shown on stage.
The pace of innovation was breathtaking, everything rolled out swiftly and assuredly.
Every announcement seemed to cater to a diverse audience.
Gemini Omni was likened to Nano Banana but designed for video content (see this strange proof).
Smart glasses are making a much-discussed return.
There are video game-like experiences that can be instantly prompted and played.
The capability for Workspace to bring documents to life with mere conversations.
A feature allowing the transformation of Google Maps images into surreal dreams seems more like a solution waiting for a problem, perhaps for Hollywood studios looking to bypass on-location shoots?
I even have Gemma on my phone, enabling in-flight conversations with a smaller model. (Thanks to American Airlines’ free Wi-Fi, I’m all set.)
And yet, the most intriguing element remains to be addressed.
Gemini and Search: Converging Evolution
Gemini is beginning to resemble Search, while Search is adopting features of Gemini.
Both platforms now include features that satisfy similar needs: keeping tabs on the web and alerting users when something of interest arises.
In Search, these are known as information agents. In Gemini, they go by Spark or Daily Brief. The connection is unmistakable.
I asked a product manager about their approach to long-term feature management and overlapping utilities. Their response was simple: “Right now, it’s all about velocity.”
Shipping fast is the mantra shared by three other product managers, all behind key I/O features initiated and deployed within this whirlwind year, 2026. It’s astounding.
The product manager elaborated, “Velocity is achieved through reduced managerial overhead.”
This implies jumping on board quickly and figuring out the finer details later.
Once You See It, You Can’t Unsee It
Armed with this understanding, the rest of the day wore a new perspective. The demos were impressive, yet I pondered: what’s the next step with these innovations?
Though I now have Gemma on my phone, one developer couldn’t provide a tangible day-to-day use case. I witnessed AI Mode’s monitoring prowess by prompting it to “keep me updated.” Despite seeing the connection of components, my questions about managing these alerts as they age went unanswered, indicating it’s still an early-stage demo.
Many features appear not to address their second-order effects thoroughly. It seems engineers are using these systems at a command line level rather than considering user interfaces.
A notable point is my current inability to delete old Gemini chats in a web browser, a functionality available in the Mac app.
Universal Cart Sparks Discussions
A frequently mentioned feature during I/O was Universal Cart, Google’s new cross-platform shopping protocol.
My opinion? If you’re Google, it’s an exciting development because, upon adoption, it further solidifies their control over the complete shopping experience. Conversely, for others, this development might be a cause for concern.
Despite these concerns, the group I conversed with didn’t seem troubled, feeling distanced from the growing anti-AI sentiment in the U.S.
Speaking with an SEO expert at a major ecommerce brand implementing Universal Cart, they related the velocity comment to their own implementation experience, describing it as feeling rushed.
Just four days before I/O, Google’s Search quality team advised publishers to “write for humans, not AI.” Shortly thereafter, the AI agent team demonstrated capabilities where Google’s own agents browse, interpret, transact, and create web content.
As Google shifts towards AI handling more tasks, the advice given to publishers starts to sound less sincere.
Impact on the Web Ecosystem
I don’t wish to undermine the engineers’ efforts. I communicated my respect for their work directly to them. Building products for search and clients myself, I can relate to frequent criticisms over compliments.
Still, the potential downside of overlapping features, difficulty in managing or reconciling data could lead to significant technical challenges later. The current AI strategy appears to be: prioritize feature utilization first, reconcile later.
Nevertheless, I admire Google’s rapid progress and look forward to future developments. Leveraging substantial resources, they can experiment comprehensively to identify successes.
Regrettably, my enlightening conversation with the product manager was abruptly concluded as we were asked to vacate the premises.
Spotting the Bright Spots
Google reports unprecedented high search query volumes. They are enhancing authentication and provenance through SynthID’s expansion into Search and Chrome, welcoming new partners like OpenAI, and integrating C2PA content credential verification.
These are indeed significant accomplishments.
However, the relentless pace might lead to unforeseen challenges. My hope is that the quest for speed doesn’t further destabilize the already-fragile web ecosystem.
In conclusion, it’s undeniably an exhilarating era for search technology.
I’ve just delved into Goodie’s enlightening AI search traffic report for early 2026, covering the period from January to April, and I’m excited to share my insights with you. This report dives into trends in usership, referral traffic, and marketing considerations, offering a comprehensive view of the shifting landscape.
You’ll want to pay particular attention to how ChatGPT’s dominance is starting to wane, with some surprising contenders like Claude and Gemini making waves. This shift could significantly impact how marketers strategize their efforts in AI-driven search optimization.
The data reveals fascinating patterns in user habits and referral traffic, which could inform future marketing strategies and the allocation of resources. For a full dive into these emerging trends and what they might mean for businesses, I encourage you to explore the detailed findings of the report.
When I attended Google Marketing Live 2026, I witnessed firsthand how Gemini is reshaping the world of Search, advertising, commerce, and measurement. The event highlighted the move towards a more conversational, AI-driven ecosystem.
This year, the focus was on agentic AI, conversational Search, automated creative production, and AI-assisted shopping. Google rolled out tools across Search, YouTube, Merchant Center, and Analytics aimed at making campaigns more autonomous, predictive, and interconnected.
Let me take you through the biggest announcements from Google Marketing Live 2026.
Google Introduces a New Generation of AI-Powered Search Ads
Google rolled out new Gemini-powered ad formats that enhance AI Mode and conversational Search experiences.
The updates include:
Conversational Discovery ads
Highlighted Answers
AI-powered Shopping ads
Business Agent for Leads
These innovative formats are crafted to be more contextual and interactive by embedding AI-generated explanations and conversational experiences directly into Search journeys.
Plus, Google expanded its Direct Offers pilot with AI-generated bundles, native checkout, and travel promotions seamlessly integrated into AI-assisted Search experiences.
Google Launches Ask Advisor Across Ads, Analytics, and Merchant Center
At the event, Google introduced Ask Advisor, a Gemini-powered AI collaborator that bridges Google Ads, Analytics, Merchant Center, and the Google Marketing Platform.
It functions as a unified assistant to help marketers:
Build campaigns
Analyze performance
Receive recommendations
Automate operational tasks
Google assures that Ask Advisor expedites the process from planning to optimization by pulling insights across platforms.
Google Upgrades Measurement with Meridian and Predictive AI Tools
Google announced new tools for measurement and forecasting within Google Analytics 360.
Meridian, an open-source marketing mix model, is being integrated directly into Analytics 360, along with Qualified Future Conversions (QFCs), a predictive reporting metric powered by Gemini.
These tools will assist advertisers in:
Improving media mix modeling
Forecasting campaign outcomes
Measuring incrementality
Linking current ad activity with future revenue signals
Have you ever wondered how AI manages to stay grounded in reality? As I delve into the fascinating world of LLM grounding, I uncover how AI models maintain their accuracy, and why this is crucial for your brand’s visibility and success across platforms like ChatGPT and Gemini.
Understanding how AI functions in this way is not just about technical curiosity; it’s about knowing how to leverage these tools to enhance your brand’s presence and credibility online. Join me as I explore the role of LLM grounding in shaping AI’s effectiveness and reliability.
