"Not a Ranking Factor": What Google, ChatGPT, and Claude Won't Tell You

Google uses the term narrowly. A ranking factor moves you up or down among the pages already competing for a query. It says nothing about:

• Whether you made that set of candidates in the first place
• Whether you qualify for a search feature at all
• Whether you get 82% more clicks from the same position

"Not a ranking factor" is almost always a denial about that one narrow thing.

Don't get caught up in the corporate semantics. Google uses a literal definition of ranking, likely to protect itself legally. But, don't misinterpret this definition as it's not important or it won't help your results. Without the things we talk about in this guide, you may never make it into the search results at all, let alone rank higher.

There's no dial in the algorithm Google turns up when your page looks authoritative. As Google puts it:

"While E-E-A-T itself isn't a specific ranking factor, using a mix of factors that can identify content with good E-E-A-T is useful."

What Google actually measures is a bundle of signals that approximate the goal:

• Clear authorship with real credentials
• Real sourcing and citations
• A reputation other sites confirm
• Evidence of firsthand experience

These are the things that move the needle.

The quality raters people obsess over work the same way. Google's explicit: "Search raters have no control over how pages rank. Rater data is not used directly in our ranking algorithms." They tell Google whether its signals are calibrated correctly. They don't set rankings.

So, the disclaimer doesn't actually change what you do. You still build the proxies. Real authors, real credentials, real sourcing, genuine firsthand experience on the page. Whether you call it E-E-A-T or not is beside the point - it's what you'd build anyway.

Three terms that get blurred together, but they are worth separating:

• Schema.org is the vocabulary, the full library of types and properties
• Structured data is that vocabulary placed on your page in machine-readable form
• Google's supported types are the subset that earns a richer result

Think of it like a nutrition label. The calories are already in the food. The label doesn't add anything; it just states what's already there in a standard format anyone can read at a glance. Schema is that label for your web page: it tells Google and the AI engines what the page already says, in a format they can parse instantly.

On this one, the denial is just true. In April 2025, Google's John Mueller said it plainly: "Structured data won't make your site rank better."

So why do it? Two reasons.

Structured data won't make your site rank better. - Google's John Mueller, April 2025

The first is eligibility for a better-looking result. The numbers aren't small. From Google's own published case studies:

• Nestlé's rich-result pages earned 82% higher click-through rate than non-rich pages
• Rotten Tomatoes measured 25% higher CTR on pages with structured data
• Food Network reported a 35% increase in visits after enabling search features

Same position. Far more clicks.

The second reason is comprehension and identity. Schema removes ambiguity that plain text can't:

• Is "Apple" the company or the fruit?
• Is 4.5 a price, a rating, or a phone number?
• Is "Dr. Smith" the author or just someone the article mentions?

It also connects your business to its verified identity across the web in a way your own page copy can't do alone. This matters more every year, because the things reading your page aren't just Google anymore. ChatGPT, Claude, Google's AI answers - they all parse pages to decide who to mention and cite. Clean markup is how you make sure they describe you correctly instead of approximately.

One rule that comes with all of it: you can't print calories that aren't in the food. Google's guidelines are clear: "don't add structured data about information that is not visible to the user, even if the information is accurate." Fake reviews, prices you don't actually charge, content the visitor can't see - Google treats that as deception and it carries a real penalty.

The first two are honest denials. This one wasn't.

For years Google publicly downplayed two ideas: that user clicks influence ranking, and that it keeps any kind of site-wide authority score. Then in May 2024, a large set of Google's internal Search documentation was accidentally published and widely reported on. Among the thousands of ranking-related features it listed:

• A click-based re-ranking system recorded under internal names like "goodClicks," "badClicks," and "lastLongestClicks"
• A "siteAuthority" signal, the thing Google had spent years saying it didn't have

Around the same time, in sworn testimony during the U.S. antitrust trial, a senior Google search executive confirmed the click system is real and important, drawing on many months of user data.

That doesn't mean schema is secretly a ranking factor. Across all those leaked features, structured data didn't appear as a ranking signal. On that specific question, the public guidance and the internal reality actually agree.

This one might be the most brazen of all of them.

At a search conference in 2023, a Google representative said links are no longer a top-three ranking factor and that people overestimate their importance. Google has pushed a version of this line for a few years now. The general message being that links are losing steam, content is what really matters, and you can rank without them.

But look at what the data actually shows.

Backlinko's study of 11.8 million Google search results found that the number one result has 3.8x more backlinks than pages ranked two through ten. Nearly four times as many. The number one result also has three times more referring domains than everything else on page one.

And then the 2024 leak happened. Among the thousands of ranking features documented in those internal Google files were not one but multiple active versions of PageRank, the link-based authority system Google retired from public view in 2016. The variants listed included RawPageRank, PageRank2, PageRank_NS, and FirstCoveragePageRank. Not legacy code. Active components of the ranking system.

So the public message is "links matter less." The internal reality is the opposite.

Why does Google do this? Probably because telling the world that backlinks are a monster signal is an invitation for manipulation, and they've spent years fighting link spam at scale. If every business understood how much links move rankings, the incentive to buy them gets stronger. Downplaying it is, at least in part, a spam-reduction strategy. But it's still not the truth.

This one runs the other direction, and the pattern is just as costly.

Google confirmed page speed as a ranking factor for desktop in 2010. They expanded it to mobile in 2018. Then in 2021 they announced Core Web Vitals with months of advance notice, a whole set of speed and user experience metrics the entire industry scrambled to prepare for. Developer time, agency budgets, tool subscriptions. It was treated as a major inflection point for search.

Then Mueller said: "We've been pretty clear that Core Web Vitals are not giant factors in ranking."

The actual impact turned out to be a tiebreaker at best, meaningful only when two pages are close to identical in content quality and authority. A slow page with great content still outranks a fast page with bad content. That was true before the update and it stayed true after.

Google didn't deny this one. They confirmed it as a factor. But the gap between how it was announced and how it actually performed cost a lot of people a lot of time and money optimizing for something that barely moved the needle. The contradiction here was an overstatement, which turns out to be just as misleading in practice.

AI Overviews are Google's own product, so the contradiction here is Google's too.

The popular belief is that AI Overviews just quote the top search result, and that AI search rewards the same organic winners, more and more over time. It feels right, but it's getting less true.

The cleanest measurement comes from Ahrefs, who ran the same study twice. In July 2025, 76% of the pages AI Overviews cited also ranked in the organic top 10. By March 2026, that was down to 38%. Cut in half in eight months.

What happened? Two things. Google's AI Overviews don't run a single search - they fan it out into a batch of related sub-searches and pull sources from all of them, so pages that never ranked for your exact query get pulled in. And in January 2026 Google swapped the underlying model to Gemini 3, which reshuffled something like 42% of the cited sources almost overnight.

So does ranking still matter? Yes, a lot. Position one is still the single strongest predictor of getting cited, and across a Zyppy large meta-analysis of citation studies, search rank came out as the number two factor overall. But "rank number one and you'll get quoted" stopped being a safe bet. The pool the AI quotes from is wider and looser than the top 10 now.

The honest version: rank is still your biggest lever, but it's no longer a lock. Getting cited has become its own game layered on top of ranking. The signals for that game lean somewhere specific, and that's exactly what Part Two is about.

When an AI answers a question, the information came from one of two places.

• Known: Baked into the model when it was trained, months or years ago. Frozen. Nothing you do today touches it.
• Found: Looked up live, right then, from the web. This is the part anyone can influence.

Every "get your brand into ChatGPT" pitch is really about the second one, the found answers. The trouble is the models don't look things up every time. A lot of the time they answer from memory, and memory is closed to you until the whole model gets retrained, which happens on a clock you don't control.

So how do you tell which kind of answer you're looking at? The citations. When a model looks something up, it shows its sources - the little links and source cards under the answer. When it answers from memory, there are none. Sources on screen means a live search happened and your content had a shot at being in it. No sources means the model answered from what it already knew, and nothing you published this week was ever in the running.

That single tell (sources or no sources) is the whole game. It tells you which questions you can actually compete for, and which ones were already decided inside a model you can't reach. The rest of this section is about winning the ones you can.

• ChatGPT searches Bing, mostly. Chapter 15 has the catch.
• Claude searches Brave.
• Google's Gemini and AI Overviews search Google.

We get into what that means for each one in Parts Three and Four. The point right now is the principle: if you're not in the index an assistant borrows, it cannot cite you, full stop. Your page can be brilliant, it can rank number one on Google, and ChatGPT still can't quote it if Bing never indexed it. People pour months into Google and never check whether Bing has even crawled them.

There's an honest catch that runs through this whole section, though. Being in the index gets you eligible, but it doesn't get you picked. Ranking well inside the index turns out to be a weak predictor of which page actually gets quoted; the assistants regularly cite pages sitting well down the results. So, step one is to be in the right index. Step two is to win a selection game that isn't really about rank at all.

Look at where the AI engines pull their cited sources from and a small set of names keeps coming up, way out of proportion to their share of the web. The freshest hard count is Ahrefs' mid-2026 study of Google's AI Overviews: YouTube and Reddit lead, each around a fifth of the citations going to the top sources, with Wikipedia further down. The lean shifts by engine:

• ChatGPT pulls more from Wikipedia
• Perplexity relies more heavily on Reddit
• Claude leans toward big-name journalism like the Times, the Atlantic, and the Economist. A couple of platforms pull ahead, but the list is long and it reshuffles constantly.

Two things about that you should know.

First, some of it is bought. Google pays Reddit a reported 60 million dollars a year for access to its content. OpenAI signed its own Reddit deal around the same time, reportedly in the 70-million range. So, when you see Reddit threads cited constantly, that isn't purely organic merit. There's a commercial arrangement underneath it. Those deals buy training and data access, not a guaranteed citation slot, and nobody has proven the money directly causes the citations.

Second, the mix moves fast. Reddit's share of citations on Google more than doubled, from about 2% to 5%, in the months to January 2026, per Tinuiti's tracking across seven engines, and the engines openly tune which sources they lean on. So, any plan built on "get cited wherever the AI loves right now" is building on sand, because what it loves is a moving target.

Then Semrush put a scale on it across 126 million prompts, and two things stand out. The engines don't even quote the same number of sources: Google's Gemini averages roughly 3 per answer; ChatGPT pulls 15. And the biggest sources get cited far more than they ever get named - what Semrush calls the Source Surplus. Wikipedia is quoted 4.3 times for every time an AI actually says the word "Wikipedia," and on Google's AI Overviews, YouTube is cited 83 million times. These are the sites AI treats as infrastructure: quoted constantly to build answers about every other brand.

The takeaway isn't "go spam Reddit." It's that the engines pull from a small, shifting, partly-paid set of big platforms, and the right move is to be the brand that gets talked about across a lot of them - not to chase whichever one is hot this month.

So, if ranking isn't the whole game and chasing the hot source is a trap, what actually moves the needle? Mostly, it is what other people say about you.

Start with mentions versus links. Google built its world on backlinks, one site vouching for another with a link. The language models learned differently. They were trained on raw human text, so what shapes how they "know" your brand is how often and how credibly you get talked about, with or without a link attached. Ahrefs studied 75,000 brands and found that plain web mentions of a brand correlated with AI visibility about three times more strongly than backlinks did. The single strongest signal they found was being mentioned on YouTube. Your number of pages, the thing SEO has chased for twenty years, barely registered.

Then there's where those mentions live. A Muck Rack analysis of millions of AI citations found that around 84% of everything the engines cite is earned media, the real third-party coverage, journalism, analyst write-ups, etc. Paid and advertorial content was a rounding error, about 0.3%. Brands turned out to be more than six times as likely to get cited through someone else's site than through their own. And when a team took the exact same content and pushed it out through real news outlets instead of leaving it on the company blog, its AI visibility jumped by a median of 239%, the citation rate going from 8% to 34% for identical words.

Brands that get mentioned a lot, get talked about on YouTube, and get covered by real publications also tend to be genuinely well-known, and you can't cleanly separate the cause from the effect. You can't game one number in isolation. But the shape is unmistakable, and it lines up. Get genuinely talked about, by other people, in real places. That's the same thing that's always built a brand. The AI engines just made it measurable.

Semrush gave the pattern a name after combing 126 million prompts: the Citation Core. In every industry, AI leans on a small, trusted set of third-party sites. Software runs on G2, Capterra and Forbes Advisor. Finance on NerdWallet, Investopedia and Bankrate. Beauty on Allure and Byrdie. The most direct path to authority is earning your place in your industry's Citation Core. That is where AI builds its answers about every brand in your space.

Patagonia is the cleanest example anyone has published. Its AI visibility held at 79 or 80 every single month, and it earned that standing through six specialist gear-review sites plus Reddit, which together account for more than 65,000 mentions of the brand in AI answers - more than all of the high-profile media and news sources combined. Patagonia built an ecosystem of other people describing it the same way, and the models read that ecosystem back.

This is why the boring old earned-media playbook(getting written about, quoted, and referenced by people who aren't you) is becoming the new AI-visibility playbook. Because that's what the models read.

A Semrush study put a number on it: about 62% of AI citations are what the researchers called ghost citations. The model used your page to build its answer, but your brand name never appears in what the user reads. Your information is in there. You aren't.

In classic search, being the source meant a blue link with your name on it and a click. In an AI answer, being the source can mean a footnote nobody opens, or nothing at all, while the AI states your information as its own. You did the work, the user gets the answer, and your name never enters their head.

It cuts both ways for strategy. Raw citation counts can flatter you. You might be "cited" constantly and still invisible as a brand. The thing worth optimizing for isn't just being used as a source, it's being named. In other words, you want to be the brand mentioned inside the answer, not just the link underneath it. Which brings you right back to the last chapter: get mentioned, in context, by sources the model trusts.

What often gets overlooked in all of this, however, is what people do next. When an AI recommends a brand, most users don't tap the source link. They open a new tab and search the name, maybe even a day or two later. Similarweb followed real user journeys for six months and measured it: 55.9% of the traffic an AI recommendation produced arrived through search, and only 8.8% came as a direct click from the AI.

Which flips how you measure the whole thing. Counting AI referral clicks tells you almost nothing, since that's just 8.8% of the story. The real signal is the lift in branded search once you start showing up in the answers. Get named by the model, and the payoff lands back in Google a few days later, under your own name. In other words, AI citations may be the reason someone clicks on your site, but it won't show up in last-click attribution.

There's a file going around called llms.txt, pitched as "the robots.txt for AI." A little text file you drop on your site to tell the AI models what you're about and which pages matter. Agencies are charging to set it up.

Here's the problem. Nothing reads it.

A developer proposed llms.txt in 2024 as a way to help coding assistants find their way around documentation. It was never a ranking or visibility tool. The SEO world repurposed it into one, and that repurposing is the whole contradiction.

Google's John Mueller compared it to the old keywords meta tag, the long-dead tag where you typed what your page was about and the search engine had no reason to believe you. In his words: "None of the AI services have said they're using it, and you can tell when you look at your server logs that they don't even check for it."

And people did check their server logs. Wislr watched 48 days of traffic and found zero requests for llms.txt from any AI crawler. OtterlyAI watched 90 days across more than 60,000 AI bot visits - the file got requested 0.1% of the time, worse than a random content page. Ahrefs looked at 137,000 domains and found that of the ones that had published an llms.txt, 97% of the files got zero requests in a month.

The best part? Google accidentally published an llms.txt on its own developer docs late last year. Mueller's entire public response was "hmmn :-/" and the file was deleted within hours.

There's a narrow, legitimate use for llms.txt - helping a coding assistant navigate a set of technical docs. As a way to get found or ranked by AI search, it does nothing. Skip it and spend the time on something a model actually reads.

A lot of site owners, worried about AI companies scraping their work, added rules to block the AI bots. That was a logical instinct, but here's what you may not know.

First, blocking the training bot doesn't remove what's already in the model. OpenAI's GPTBot collects content to train future models. Block it today and you might affect the next training run, but you do nothing about everything already learned. There is no un-training.

Second, blocking the training bot does nothing to stop you being cited in live answers, because training and live lookups use different bots. A BuzzStream study of 4 million AI citations found 88% of the sites blocking OpenAI's training bot got cited anyway, because the citation came from the live-search bot, not the training one.

Then there's this trap with Google. Google offers a token called Google-Extended that opts you out of training its Gemini models. It sounds like the lever for staying out of AI Overviews. It isn't. In Google's own words, "Google-Extended doesn't impact a site's inclusion or ranking in Google Search." AI Overviews are built from the normal Google index. The only way to keep your content out of AI Overviews is to block Google Search entirely, which means vanishing from Google. Nobody wants that trade. In that same 4-million-citation study, 92% of the sites blocking Google-Extended showed up in AI citations anyway.

So blocking is mostly theater unless you're willing to go nuclear and leave Google search altogether. Decide what you actually want before you touch the robots file.

"Optimize for AI" usually gets shortened to "ChatGPT runs on Bing." A year ago, that was close enough. The real picture, from the best look anyone's published inside the thing, is stranger.

The cleanest data comes from a 2026 study by Ali San Kaya at Maestra, who ran 79 prompts across ChatGPT Plus, ChatGPT Business, and Gemini, three times each, over three model versions this year, and captured 370,000 search results and 13,000 citations behind the answers, all scoped to one product category he knows cold: AI speech and translation tools. The newer models started exposing, right in the raw response, which search provider returned each cited link.

Here's what it shows:

• ChatGPT Business grounds its answers in Bing. About 95% of its cited links come straight from Bing's index.
• ChatGPT Plus grounds its answers in Google - specifically a scraped copy of Google's results, pulled through a third-party scraper (Bright Data, whose whole business is scraping Google). About 94% of Plus citations come from that scraped Google, not Bing.
• And 12 to 15% of citations on both tiers don't come from a search engine at all. They arrive through an internal OpenAI channel that leans on Wikipedia, arXiv, and major tech press - sources you can't reach by ranking anywhere.

So, the exact same question, asked of ChatGPT, gets answered off Bing or off a scraped Google depending on whether the person is paying for Business or Plus. The two tiers are, underneath, two different search engines wearing the same chat box. If you only ever check your Bing presence because "ChatGPT is Bing," you're invisible to every Plus user, who's actually being served Google.

The practical read is almost annoyingly old-fashioned. You need to be in Bing and ranking in Google, because depending on the user's plan ChatGPT is reading one or the other. And that internal Wikipedia-and-press channel is just the earned-media game showing up again, in a spot no amount of on-site SEO can touch.

ChatGPT also doesn't run your question once. It fans it out into several reworded searches and pulls from all of them, so the page it cites was often ranking for how an external source described you, not how you present yourself on your own website.

Remember the known-versus-found split we talked about earlier? ChatGPT is where you can actually watch it happen, because people have measured it.

ChatGPT decides whether to search the web based on what the user seems to want, and the gap is enormous. When someone's just learning, like "what is a CRM," ChatGPT answers from memory and only bothers to search around 10% of the time. When someone's clearly choosing, like "best CRM for a small clinic," it searches the web more than 70% of the time. Choosing triggers a lookup. Learning doesn't.

That should change where you spend effort. The questions where ChatGPT goes and looks, and where your content can actually land, are the high-intent ones: best, top, compare, alternatives, near me, the fresh stuff that changes. The encyclopedia questions are mostly answered from a frozen memory you can't reach. So if you're trying to show up in ChatGPT, don't fight for the definitions. Fight for the moments someone's deciding, because those are the moments it actually goes looking.

Broader survey data lines up with this. When Similarweb asked people which tool was most useful at each stage of a decision, AI ran away with discovery: 35% named it against 13.6% for search. By the time they were choosing where to buy, search had nearly caught back up, 22.1% to AI's 24.3%. AI owns the top of the funnel, where nobody has typed a query yet. Search still owns the bottom, where the money changes hands. The same content that earns the AI answer is what ranks for the purchase-intent search sitting right under it.

When Claude searches the web, it's running your query through Brave Search, a smaller independent engine most people have never used. It came to light when Simon Willison noticed Brave appearing on Anthropic's list of subprocessors and found a setting literally named for Brave buried in Claude's web-search code, then confirmed it by matching Claude's citations to Brave's results one for one.

Why does this matter? Because Brave's index isn't Bing's and isn't Google's. The pages that surface in Claude can be different from the ones that surface in ChatGPT or Google, for the very same question. Showing up in Claude is its own game, played on an index almost nobody is optimizing for. That cuts both ways. There's less competition, but you have to actually be in there.

And Claude's audience isn't a rounding error. It skews toward developers, operators, and professionals, the people who research a purchase carefully and sign off on real budgets. Claude is also the most reluctant of the big assistants to cite at all. It pulls sources on only about half its answers, but when it does, it tends to cite more of them and lean on serious publications. For a lot of businesses, being the brand Claude names in that room is worth more than another impression somewhere else.