Updated 19 October 2026 · 9 min read · By Matthew Bertram
AI SEO is search engine optimization with one more target: the answers that AI engines write. Most of the work is ordinary SEO done well. The new part is that an AI answer credits a few sources instead of listing ten links, so each section of a page has to answer its own question clearly enough to be quoted and credited on its own.
AI visibility is the outcome you track: whether AI answers bring up your business, and whether they get it right. AI SEO is the work that moves it.
What stays the same?
Most of the job. Google's AI optimization guide says "The best practices for SEO continue to be relevant because our generative AI features on Google Search are rooted in our core Search ranking and quality systems." It goes further: "From Google Search's perspective, optimizing for generative AI search is optimizing for the search experience, and thus still SEO."
Read that as Google's on-record position about its own product. AI Overviews and AI Mode draw on Google's index. The work that gets a page indexed and ranked there also makes it eligible for those answers. None of it tells you how ChatGPT or Perplexity pick their sources.
The parts that carry over are the ones that were always hard. A page has to be crawlable and indexed. It has to answer the question better than the pages around it. The site has to be known for its subject, which SEOs call topical authority. Other sites have to talk about you, too: our guide to AEO covers an Ahrefs study in which brand mentions around the web tracked AI Overview visibility far more closely than backlinks did. No AI tactic makes up for a page Google can't index.
What is actually new?
The finish line moved. A results page lists links and lets the searcher choose. An AI answer makes that choice for them. It writes the reply itself and credits a few sources.
Each section has to stand on its own. An engine that credits your page is drawing on a passage, not your whole argument, so if the answer to a heading sits three paragraphs below it, that passage is hard to use. Put the answer in the first lines under the heading.
You can be cited without ranking for the question. Google says its AI features can run extra searches on related subtopics before answering, which it calls query fan-out. A page can be pulled in for one piece of a question without ranking for what the person typed, and our guide to how AI Overviews choose their sources traces how that happens.
More pages won't cover the gaps. It's tempting to publish a separate page for every way people phrase a question, including the related searches an AI runs. Google's guide names that exact move as a problem: doing it "primarily to manipulate rankings or generative AI responses in Google Search violates Google's scaled content abuse spam policy." One strong page that answers the main question and its follow-ups is the safer build.
A bolt-on FAQ doesn't rescue a page. One shortcut we keep running into is a block of questions stapled to the bottom of a page whose body answers none of them, when those questions belong in the body as headings with the answer directly beneath.
Do you need to optimize separately for ChatGPT and Google?
For the writing, no. A page that answers buyer questions clearly, with a source for every number, reads the same to every engine, though two things do differ: how each engine reaches your site, and how you check the result.
Reach first. AI Overviews and AI Mode answer from Google's index. Googlebot has to crawl the page before either can use it. OpenAI finds pages for ChatGPT's search features with its own crawler called OAI-SearchBot, and its crawler documentation says sites that opt out of it won't be shown in ChatGPT search answers. Perplexity runs PerplexityBot for the same job on its side. A firewall rule or a hosting default can shut out those two crawlers while Googlebot gets through, and nobody notices until they check. Our guide to blocking AI crawlers lists which bot does what.
Checking differs too, and the last section of this guide covers it.
Which AI SEO tactics have evidence behind them?
Fewer than the pitches suggest. We rated nine tactics that AI SEO checklists push, using five labels:
- Lab support: a controlled experiment found a gain, under conditions unlike today's engines.
- Correlation only: it shows up alongside AI visibility, but nobody has shown it causes citations.
- Unproven: we found no controlled test either way.
- No gain found: researchers measured it and saw no lift.
- Retired: the search feature it was built for is gone.
| Tactic | Evidence | Our call |
|---|---|---|
| A direct answer under every heading | Unproven | Do it. It helps human readers whatever the engines do |
| Statistics with a linked source | Lab support | Do it. The GEO paper found gains, but GPT-3.5 wrote the answers in its main test |
| Mentions of your brand on other sites | Correlation only | Earn them. This row moves slowest |
| A named author with credentials | Unproven | Do it, with a real bio page. See E-E-A-T |
| A visible last-updated date | Unproven | Show it. Change it only when the page changes |
| Comparison tables | Unproven | Use one where buyers weigh two options |
| Schema markup, tested on pages already heavily cited | No gain found | Use it to say who you are, not to win citations |
| FAQPage markup | Retired | Don't add it to new pages |
| An llms.txt file | No gain found | Optional. See below |
Two rows need a word. Ahrefs tested schema against matched control pages and found no citation gain from adding it to pages that were already heavily cited. Our guide to schema and AI citations covers what markup is still good for. FAQPage markup lost its main job when Google stopped showing FAQ rich results, a change its documentation dates to May 2026.
Look at the Unproven rows we still recommend. They're on our list because each is cheap and helps the person reading the page, not because a test showed it wins citations. Hold any pitch to the same line. When someone sells one of these as proven, ask to see the test, including the control it was measured against.
Is llms.txt worth adding?
Not for AI search. An llms.txt file is a plain-text guide to your site, placed at its root for AI tools to read. Our guide to blocking AI crawlers explains what the file can't do. The open question is whether anything reads it.
Ahrefs' June 2026 study says very little does. The search crawlers behind ChatGPT and Perplexity, plus Claude's, "made only a couple of hundred fetches across thousands of sites" between them. Among AI tools, Ahrefs puts GPTBot first. That's OpenAI's training crawler. Claude-Code, Anthropic's coding agent, came second, not counting an agent called statespace-indexer whose IP ranges Ahrefs couldn't confirm. Ahrefs also found that AI bots never requested a file that didn't exist: "They never go looking." So a business that publishes its first llms.txt file hoping ChatGPT or Perplexity will notice it shouldn't count on that, because most of the requests these files receive come from bots and tools that aren't gathering sources for a search answer.
SE Ranking came at it from the citation side. It looked at nearly 300,000 domains and found no link between having the file and AI citations. Its write-up adds: "Removing this variable from our XGBoost model actually improved its accuracy." XGBoost was the prediction model it built. The file didn't just fail to predict citations. It made the predictions worse.
Google's position matches. Its guide says Search ignores llms.txt files. It also says it's "completely fine" to maintain one "for other services or systems that use these files." Google's John Mueller made a related point on the Search Off the Record podcast, as Search Engine Journal reported: a file a site writes about itself, he argued, can't help a language model choose between sites. He carved out one use: "If someone is already on your website, maybe some kind of automated system is helpful." His example was an AI agent trying to buy a photograph from a specific site.
Google's developer tools treat the file as an agent feature instead of a search one: Lighthouse 13.3, Google's page-audit tool, added an experimental Agentic Browsing category that checks whether a site provides the file.
There's a cost to getting it wrong. Ahrefs warns: "Agents are built to trust this file, and potential bad actors are already probing llms.txt for prompt injection." Prompt injection means planting instructions in text an AI agent reads, so the agent acts against its own user. Ahrefs' summary: "A stale or compromised file misleads every agent that reads it." Markdown copies of your pages, built for bots, belong in the same drawer. Google's guide says Search doesn't use them, and Mueller has called building them "a stupid idea."
Our call: publish a short, accurate llms.txt only if people point AI agents or coding tools at your site, such as a software company with developer documentation. Check it whenever the site changes. Don't pay anyone to write one, and don't count it as AI search work.
Where should you start?
Start small, in this order.
- Make sure the engines can reach you. Check that your robots.txt file and firewall let Googlebot in. Then check the same for OAI-SearchBot and PerplexityBot. Search Console's page indexing report shows whether Google has indexed the pages you care about.
- Rewrite the pages behind five to ten buyer questions. Pick the questions people ask before they contact a business like yours. Put the answer in the first lines under each heading, and link a source for every number.
- Take a reading before you change anything. Ask those questions on each engine several times and log who gets cited. Make one change. Once the page has been recrawled, ask again the same way.
Step three gets skipped most. Without a before reading, a citation you spot next month proves nothing, because you can't tell whether it was there all along.
How do you know if it worked?
Two engines report on it. Google's generative AI performance report in Search Console counts how often links to your site were shown in AI Overviews and AI Mode. It breaks that count down by page and country, with no view of the searches behind it. Bing Webmaster Tools has an AI Performance report that counts citations across Copilot and Bing's AI summaries, and it does show the searches, which Microsoft calls grounding queries, though Microsoft describes the data as a sample.
ChatGPT and Perplexity give site owners no citation report that we've found. You check them by hand: the same questions, asked the same way, several times each. Our guide to checking AI citations has the full method.
The repeats matter because AI answers change from one asking to the next. In our team's variance study at EWR Digital, a visibility figure built from one run per question carried a 95% confidence interval of roughly plus or minus 33 points. The study's verdict on that number: "It is not a measurement." Seven or eight runs per question tighten the interval to about 10 points either way.
Analytics covers the other half: people who click through from an AI answer to your site. Brie Anderson walked through how to attribute that traffic on an episode of this show.
So when anyone reports that your AI visibility went up, us included, ask two questions before you believe it. What was the reading before the change? How many runs per question sit behind each number?
Want a straight answer on which AI SEO work on your site is worth keeping, checked against your buyers' real questions? Book a discovery call with EWR Digital →
Keep reading: What Does AI Visibility Mean for a Business? · How Do Google's AI Overviews Decide What to Cite? · Does Schema Markup Help You Get Cited by AI?
Matthew Bertram hosts The Best SEO Podcast and is owner and CEO of EWR Digital, the SEO and AI-visibility agency behind the show. Last reviewed 19 October 2026. More about Matthew Bertram.
Run your own before-and-after test of any tactic in the table above? Send us the method and the numbers. If it holds up, we'll update the verdict and credit you.