How to Improve AI Search Visibility: The Rules Google Actually Published
In May 2026 Google finally wrote down how its AI features pick sources. Almost none of it matches what the industry has been selling. Here is what the documentation says, what it tells you to ignore, and how I check whether a page is actually showing up.
The short version
- There is no separate AI algorithm to game. Google's own documentation says there are no additional requirements and no special optimizations necessary to appear in AI Overviews or AI Mode.
- Google names the hacks it wants you to skip. Content chunking, llms.txt files and paid-for brand mentions are all called out by name in the May 2026 guide.
- The thing it spends the most words on is content nobody else could have written. Google calls it non-commodity content. It is a writing problem, not a plugin.
- You can finally measure this. Search Console added a Generative AI performance report on June 3, 2026 — impressions only, no clicks and no query data.
- Search Console cannot see ChatGPT, Claude or Perplexity. Those arrive as referral traffic in GA4, and you have to look at both.
What AI search visibility actually means
AI search visibility is how often your pages get pulled into a generated answer and cited as a source, instead of appearing as a blue link the searcher has to click.
That is the whole definition. It is a different outcome from a ranking position, but it is produced by the same work, and the sooner a business owner understands that, the less money they waste.
I have been doing this since 1990. I have watched the industry invent a new specialism every time the results page changes shape — and every time, the new specialism turns out to be the old craft with a fresh invoice attached. This one is no different, with one important exception: for about eighteen months there was no primary source at all. Everybody was guessing, loudly, and charging for the guess.
That ended in May 2026.
Google finally wrote it down
Google published a document called Optimizing your website for generative AI features on Google Search, filed under the fundamentals section of Search Central and announced on the Search Central blog. Read it yourself. It takes about fifteen minutes and it will save you from most of the sales calls you are getting.
The companion page, AI features and your website, is even blunter. It says there are no additional requirements to appear in AI Overviews or AI Mode, and no other special optimizations necessary.
I want to be careful here, because that sentence gets misquoted in both directions. It does not mean AI search is irrelevant. It means there is no second door. The pages that earn ordinary organic visibility are the pages that get quoted, so the work is the work you already should have been doing.
The one idea worth the whole document
Google keeps returning to what it calls non-commodity content — material that could not be produced by somebody who had only a search engine and an afternoon. Its own example contrasts a summary of existing advice against a first-hand account of a real job, right down to what was found inside the sewer line.
That is not a technical instruction. It is an editorial one, and it is the reason a one-person shop can outrank a content farm.
What Google tells you to skip
This is the part I did not expect. Google's guidance names specific tactics and tells site owners not to bother with them. I have had four separate clients quoted for at least one of these in the past year.
| Tactic | How it gets sold | What the guidance says |
|---|---|---|
| An llms.txt file | "Hand the AI a menu of your best content" | Named by name as a tactic to skip. No major system has committed to reading it. |
| Content chunking | "Reformat every page so an AI can lift passages" | Also named as a tactic to skip. Write for people, in the order people read. |
| Paid mentions on "best of" listicles | "Seed mentions so the models learn your brand" | Called out as inauthentic mentions. Earn the mention or skip it. |
| Schema markup as a citation lever | "Structured data makes the AI quote you" | No special markup is required. Markup still has to match the visible text. |
| An "AI rank tracker" reporting position 3 | "Track your AI ranking week to week" | Google's only first-party data is impressions. There is no position and no query. |
| First-hand, non-commodity content | Rarely sold at all — it is slow and it cannot be automated | This is what the document spends the most words on. |
There is a pattern in that table. Everything Google tells you to skip is cheap, fast and sellable. The thing it tells you to do is slow, specific and impossible to outsource to a tool. That is not a coincidence, and it is roughly the same shape as every SEO fashion I have watched come and go since the first web directories.
How I check whether a page is actually showing up
Until June 2026 there was no honest answer to "am I in AI results?" People took screenshots of ChatGPT and treated them as data. There is now something better, though it is narrower than most people expect.
1. The Search Console generative report
Google launched the Generative AI performance report on June 3, 2026. It sits under Performance, next to the standard Search results view, and counts an impression whenever one of your URLs appears in AI Overviews, AI Mode or AI features in Discover.
What you get: impressions by page, country, device and date. What you do not get, at least for now: clicks, click-through rate, or the query text. The Search Console help page also notes it is still rolling out, so a property can be perfectly healthy and simply not have the tab yet.
2. A GA4 referral segment
Search Console only sees Google. A customer who asks Claude or Perplexity for a Tucson contractor and then clicks through arrives at your site as plain referral traffic, under the referring domain, in Google Analytics 4.
Build the segment once and look at it monthly. On most of the local sites I manage the volume is still small next to organic search — but it converts unusually well, because the person arrives having already been told you are the answer.
3. Prompt testing, treated as anecdote
I still ask the assistants directly, in the words a customer would use. Not "best HVAC Tucson" but "my AC quit and it is 108 degrees, who do I call in Tucson today."
But I write those results down as anecdotes, never as rankings. Ask the same prompt twice in one afternoon and you can get two different businesses. One screenshot naming your competitor proves nothing. One screenshot naming you proves exactly as little.
What I actually change on a page
Strip away the vocabulary and the list is unglamorous. This is the sequence I work through, and it is the same sequence whether the target is an AI answer or position four.
- Answer the question in the first sentence. Not in the third paragraph after a windup. If a machine has to read 200 words to find your answer, so does a person, and both of them leave.
- Write one page per question, not per keyword. This is where reading search intent properly pays for itself, because a page trying to serve two intents answers neither cleanly enough to be quoted.
- Put something on the page that only you could have written. A real job, a real number, a real thing that went wrong. This is the non-commodity test, and most service pages fail it.
- Name the author and mean it. A page attributed to "the team" carries no experience signal at all. Mine says David Cragg, links to a bio with verifiable history, and stays consistent sitewide.
- Keep the markup honest. Structured data should describe what is visibly on the page and nothing else. Contradictory schema is worse than none — more on that in technical SEO.
- Make it crawlable and fast on a phone. Google's guidance spends most of its technical section here. It is the floor, not the ceiling, but you cannot be cited from a page that cannot be fetched.
Notice what is absent: no new file in the site root, no reformatting every paragraph into bullet fragments, no buying mentions. I have tried the shortcuts on my own site over the years, which is the only place I am willing to test something I would not yet bill a client for. The shortcuts have never once outlived the algorithm that made them work.
What this means for a local business in Tucson
Local is where AI answers bite hardest, because the generated response often names two or three businesses and stops. There is no page two to scroll to.
The good news is that the inputs are ones you already control. Google's AI surfaces draw on the same local corpus the map results do, so a complete Google Business Profile and consistent listing data across the web are doing double duty. The mechanics of how the three map results get chosen belong on my Tucson SEO services page, and how to build a steady review flow has a page of its own. I am not going to repeat either here.
What I will say is the part specific to AI answers: the language your customers use in reviews becomes part of how these systems describe you. A review that says "replaced our condenser in Marana the same afternoon" is more useful to an answer engine than five that say "great service." You cannot script reviews and you should not try. You can ask better questions when you request one.
If none of your service pages currently say anything a competitor could not have said, that is the place to start — not with a new file in your site root. My write-up on why most small business sites are not ready for AI-driven search covers how common that gap still is, and industry-specific SEO covers how the answer differs by trade.
What I got wrong early
In 2025 I spent real time on formatting theories. I broke pages into short labelled blocks because that was supposed to make passages easier to lift. I added markup to pages that did not need it. On one client site I shipped an llms.txt file, because it cost twenty minutes and the downside looked like zero.
None of it moved anything I could measure, and I could not measure much, because the Search Console report did not exist yet. When Google published the guidance a year later and named two of those three things as tactics to skip, it was not satisfying. It was a reminder that in the absence of a primary source, I had been reasoning from other people's blog posts, and so had everybody else.
The pages that did move were the boring ones: the ones where I sat with an owner for an hour and got him to tell me what actually happens on a job, then wrote that down. That has been true for thirty-six years and it stayed true when the results page grew a paragraph at the top.
Who wrote this
I am David Cragg. I have worked in search and internet marketing since 1990, hold a BA in Economics from UCLA and an MBA in Marketing from the University of Washington, and I built and sold two internet marketing companies, Lotus411 and MSD2D, before going back to consulting.
Today I work with small businesses across Tucson and Southern Arizona. One consultant, $1,000 a month, no contract. Everything above is what I do on my own site first. More on my background, how the engagement works, and what it has done for other people.
I use AI tools openly in drafting, the same as I use a spellchecker. The judgment, the client history and the mistakes above are mine.
Want to know if you are showing up?
I will look at your Search Console generative report, your GA4 referrals and your top service pages, and tell you plainly whether the problem is AI visibility or something much more ordinary. Usually it is the ordinary thing.
Questions about AI search visibility
What is AI search visibility?
AI search visibility is how often your pages get pulled into an AI-generated answer and cited as a source, rather than listed as a blue link. It is a different outcome from a ranking position, but it comes from the same work. Google's documentation says there are no additional requirements to appear in AI Overviews or AI Mode, and no special optimizations necessary. The pages that earn ordinary organic visibility are the pages that get quoted. So AI search visibility is best understood as a second reward for solid SEO, not a separate product you buy on the side.
Does Google publish official guidance on optimizing for AI search?
Yes. In May 2026 Google published a guide called "Optimizing your website for generative AI features on Google Search," filed under the fundamentals section of Search Central. It is the first primary source on the subject and it is short enough to read in one sitting. The headline points are that ordinary SEO practice still applies, that unique non-commodity content matters more than format tricks, and that crawlability and page experience are the technical floor. If someone is selling you an AI search program, ask whether it matches that document. Most of what I have been shown does not.
Do I need special schema markup to appear in AI Overviews or AI Mode?
No. Google states plainly that there are no additional requirements and no special optimizations necessary to appear in AI Overviews or AI Mode. Structured data is still worth shipping, because it removes ambiguity about who you are, what a page contains, and how the entities on your site relate to each other, and because it has to match the visible text on the page. What it is not is a switch that turns AI citations on. Anyone selling schema markup as an AI visibility guarantee is selling something Google has already said does not exist.
Is an llms.txt file worth adding to my site?
Not for Google. Google's generative AI guidance names llms.txt specifically as a tactic to skip, alongside content chunking and chasing inauthentic brand mentions. No major AI system has committed to reading the file. It costs almost nothing to add, which is why it spread so fast, but a file that nobody has agreed to consume is not a strategy. I would rather spend that hour making one page answer a real customer question better than any competing page in Tucson answers it.
How do I measure whether my pages appear in AI answers?
Google Search Console added a Generative AI performance report on June 3, 2026, and it is the only first-party measurement that exists. It sits beside the standard Search results view under Performance, and it counts an impression each time one of your URLs appears inside AI Overviews, AI Mode, or AI features in Discover. It breaks impressions down by page, country, device and date. It does not report clicks, click-through rate, or query text, and it is still rolling out, so not every property has the tab yet.
Does Google Search Console show traffic from ChatGPT, Claude, or Perplexity?
No. Search Console only measures Google Search. Visits that begin in ChatGPT, Claude, Perplexity, Copilot, or the Gemini app arrive at your site as ordinary referral traffic and show up in Google Analytics 4 under the referring domain. I check both numbers: the Search Console generative report for Google's own AI surfaces, and a GA4 referral segment for everything else. Neither figure is complete, and anyone quoting you one tidy AI visibility percentage is estimating.
What does Google mean by non-commodity content?
Non-commodity content is material that could not be produced by someone who had only a search engine and an afternoon. Google's guide draws the line between a summary of advice that already exists elsewhere and a first-hand account of something you actually did. A page titled "five benefits of local SEO" is commodity content. A page showing what happened to one Tucson contractor's call volume after his service pages were rewritten is not. That distinction is the most useful thing in the document, and it is a writing problem rather than a technical one.
Can I stop Google's AI features from using my content?
Google is testing an opt-out control in Search Console that keeps a site out of AI Overviews and AI Mode. Google has said that sites which opt out will not receive traffic or impressions from those features, and that the setting is not used as a ranking signal for ordinary search results. So it is a trade, not a penalty. For a small local business I have not yet seen a case where opting out made sense. For a publisher whose revenue depends on the click, the math is different.
How often do AI answers change?
Far more often than a blue-link ranking does. The same prompt asked twice in one afternoon can name different businesses, because these systems retrieve fresh sources and generate a new answer every time. That means one screenshot of ChatGPT naming your competitor proves very little, and one screenshot naming you proves just as little. I look at Search Console impressions across a month rather than at anecdotes, and I ignore any tool that reports an AI ranking as a single fixed number.
Primary sources
Everything above is checkable. These are the documents it rests on, all of them first-party:
- Optimizing your website for generative AI features on Google Search — Google Search Central, May 2026
- AI features and your website — Google Search Central
- Introducing Search Generative AI performance reports in Search Console — June 3, 2026
- Generative AI performance report — Search Console Help
- Top ways to ensure your content performs well in Google's AI experiences — May 2025
Written by David Cragg, Tucson SEO. Published May 23, 2026. Last reviewed August 14, 2026 against Google's current documentation. AI search behaviour changes faster than this page does — if a date above has aged badly, call me and I will tell you what moved.
