We Scored 350 Top-Ranking Pages for AI Signals Across 35 SERPs. Here's What Actually Ranks in 2026
A lot of SEO advice about AI-generated content rests on one comforting assumption: Google rewards originality, punishes AI, and the best content wins. The assumption is half wrong.
We wanted to see for ourselves. So we picked 35 search queries across seven intent categories, pulled the top-ranking URLs for each, and scored every page on eight different AI-content signals. That gave us roughly 350 ranking URLs directly scored, inside a broader pool of about 500 URLs once "people also ask" links, related searches, and secondary carousel items were included. The dataset covers definitional queries, "best X" listicles, how-to guides, trend pieces, comparative queries, niche SaaS topics, and professional services.
What we found is less comforting than the standard narrative, but more useful if you want to rank. AI-generated content is already ranking in the top 10 for almost every category. The correlation between AI signals and rank is weak, just as recent research suggests. Within that weak average, specific patterns separate content that wins the click from content that just fills a slot. Here is the scoring metric we used to test the blogs: Scoring sheet.
How We Ran the Study
Before the findings, the method. We wanted to pressure-test the popular claim that high-quality, original content ranks better than AI-generated content, using a repeatable method any SEO team can run themselves. This study sits alongside the broader GEO framework for AI search, which is the strategic frame we use to think about originality across traditional Google and AI engines.
We chose 35 queries across seven buckets. The definitional bucket covered "what is" queries for marketing automation, zero trust security, vector databases, API gateways, product-led growth, customer success, DevOps, and account-based marketing. The commercial comparison bucket covered "best" queries for email marketing software, CRM for small business, project management tools, password managers, AI video generators, SEO tools, landing page builders, and CRM software. The procedural how-to bucket covered reducing customer churn, writing case studies, running SaaS demos, building content pillars, structuring sales commission plans, improving conversion rate, writing product launch emails, and building sales funnels. The trend listicle bucket covered content marketing trends, SaaS growth strategies, B2B lead gen tactics, remote team management tips, LinkedIn marketing strategies, email marketing best practices, and digital marketing trends. The comparative judgment bucket covered benefits of marketing automation, remote work pros and cons, SEO vs SEM, Slack vs Teams, HubSpot vs Salesforce, and Google Ads vs Facebook Ads. The niche SaaS bucket covered SaaS churn reduction, PQL frameworks, usage-based pricing, outbound cadence templates, SaaS onboarding checklists, RevOps frameworks, and customer health scoring. The professional services bucket covered pricing consulting services, managed IT services pricing, fractional CFO decisions, law firm marketing strategies, accounting firm growth, and proposal writing.
For each query, we fetched the top-ranking URLs in Google and scored each page on eight AI-content dimensions: lexical cliché density, triplet verb parallelism, generic opener patterns, hedging filler, pseudo-comprehensive enumeration, balanced framing tics, recursive definition, and concrete-versus-generic claims. Each dimension ran 0 to 3 for a maximum AI score of 24. Pages scoring 0 to 8 looked human-written or heavily edited, 9 to 15 looked AI-assisted, and 16 to 24 looked AI-generated with light review at most.
We did not use any detection tool. We scored what was visible in the text because classifiers are noisy and the tells we cared about, structural and lexical and stylistic, are visible to any reader who knows what to look for. The scoring sheet and full domain-by-domain breakdown is available on request.
Finding #1: AI-Generated Content Is Already Ranking Everywhere
The first finding is the uncomfortable one. Out of around 350 top-10 URLs we scored, 42 percent showed moderate to heavy AI signals (score 12 or above). About 11 percent looked fully AI-generated with minimal human editing. That number is low only if you expected Google to be filtering everything that looks AI-written.
Three patterns jumped out immediately. The average top-10 page sits around a 7.2 AI score, not a 2. The top ranking slot for most queries goes to a large-brand page like HubSpot, Salesforce, IBM, Cloudflare, or AWS that is clearly AI-assisted in places but heavily human-edited elsewhere. The middle of page one, specifically positions 4 through 8, almost always contains at least one clearly AI-generated page from a smaller publisher, and those pages stay on page one despite their tells. AI-generated content is not struggling to break in. It already lives there.
On "what is marketing automation" the top three, HubSpot at position 1, Salesforce at 2, and IBM at 3, scored between 9 and 13, while Ironhack at position 4 scored 20 with textbook triplet constructions, cliché stacking, and generic openers. Ironhack still sits on page one. On "what is customer success," Zams at position 9 scored 16 with openers like "in today's competitive business landscape," "customer-centric culture," and "trusted advisors" stacked across the same page. On "how to build a sales funnel," MNTN at position 1 scored 10 with the classic "in today's fragmented media landscape" opener, and that page still took the top slot.
The uncomfortable version of this truth is that per a recent industry survey, 97 percent of content marketing programs now use AI in their workflows, and it shows in the SERP.
Finding #2: The Correlation Between AI Signals and Rank Is Weak
The second finding confirms what Reza Moaiandin's team at SALT found independently. Rank cannot be reliably predicted from AI signal alone.
Across our 35 SERPs, the lowest-AI-signal page on the page was the top-ranking page only 31 percent of the time. In about 40 percent of SERPs, the lowest-AI-signal page ranked outside the top three. Spearman rank correlation between AI score and rank position sat at approximately -0.18, a weak negative. Less AI-looking content ranked slightly better, but only slightly, and never reliably. To put it another way, if a team optimizes for "originality" alone and ignores everything else, that team is optimizing the wrong thing about 70 percent of the time.
Most of the ranking variance is explained by factors AI-signal scoring cannot see, which is the domain authority, link graph, topic authority, and query-match quality that Google still weights heavily. For teams working on ranking in AI search in addition to Google, our guide to structured data for AI search covers why schema markup is doing more of the ranking work than copy originality in many current SERPs.
If AI signal is not driving rank, something else is. That is where the query-category split gets interesting.
Finding #3: Some Query Types Are AI-Tolerant, Others Are Not
The single most useful pattern we found is that AI-signal tolerance varies sharply by query type. SEO teams who treat all queries the same are optimizing blind.
Definitional queries are the most AI-tolerant
Queries like "what is marketing automation" and "what is account based marketing" tolerated the highest AI signals in top-ranking content. Users want a factual answer, not a point of view, so Google seems willing to surface AI-written definitions as long as they are accurate. The caveat here is that the top-ranking slot almost always went to an enterprise brand whose name carries its own authority, so originality mattered less than the domain. Average AI score for the top three in this category hit 6.9, the highest of any bucket.
Technical-definitional queries are the LEAST AI-tolerant
The interesting subcategory is technical-definitional queries for developer concepts. Zero trust security, vector databases, API gateways, and DevOps all averaged AI scores between 4 and 6 in the top three, the lowest of any category. Wikipedia and academic papers from NCBI ranked in the top seven for DevOps and zero trust respectively. The pattern is that Google seems to weight technical precision much more heavily when the topic requires it. Template language trips alarms when the audience is developers. The top three for "what is an API gateway" averaged 4.7, with F5 at 3, GeeksforGeeks at 5, and IBM at 6 holding the top positions through actual technical substance rather than marketing polish.
Commercial "best X" queries are the MOST AI-hostile
"Best email marketing software," "best SEO tools," and "best project management tools" consistently punished AI content and rewarded pages written by named operators with verifiable proof. The top three for "best email marketing software" averaged 2.3, with MarketerMilk in the lead at score 2, describing "I've used Brevo since 2016, sent 640,199 emails." Morningscore sat just behind with "350 hours of personal testing." Self Made Millennials reported "$40,000 from SEO in 2025, met the Search Atlas team in person at SEO IRL Toronto Fall 2025." EmailToolTester's reviewer narrates their own workflow. The pattern repeats across "best SEO tools," where MarketerMilk's author opens with "I led SEO at Webflow ($4.2 billion company) and then built my blog to +150,000 visitors per month," and across "best password manager," where Zapier's reviewer writes "as a tech journalist, I've been covering password security for more than a decade. I wrote my dissertation for my BSc in Psychology on the underlying reasons people can't recognize secure passwords."
These are not stylistic differences. They are verifiable proof signals that operator reviewers include and AI-written listicles cannot fake. In commercial "best X" queries, operator voice is not a nice-to-have, it is the entry fee to the top three.
Trend listicles and "how-to" pieces land in the middle
Remote team management tips, SaaS growth strategies, content marketing trends, and how-to pieces showed moderate AI tolerance. Top-ranking content in these queries rewarded proprietary data more than anything else. Siege Media's 353-marketer survey won rank six for "content marketing trends" over pieces with stronger links. Paddle's VC data on SaaS growth trends beats lighter-linked roundups. Mixpanel's 12,000-company product benchmark report outperforms plain listicles. Named-customer case study numbers do the same work. HubX describing "$106,000 recovered in 3 months, 63 percent retention" for a specific customer beats generic churn advice. The pattern is that trend and how-to content wins when it brings something to the table that a competitor cannot synthesize from public data.
Comparative and professional services queries reward credentials
"HubSpot vs Salesforce" top ranks were a vendor product page showdown (both scored 10 to 11 on AI signal, heavy marketing language), but the third slot went to RevOps Co-op, an admin who wrote "I've been the primary CRM for Salesforce, HubSpot, Zoho CRM" and walked through the actual trade-offs with a score of 2. Professional services queries like "how to price consulting services" sent top ranks to HBR where articles from Dorie Clark, Alisa Cohn, and Marshall Goldsmith, all verifiable named authorities with books and consulting practices, held the top positions. Named credentials carry professional services content the way proprietary data carries listicles.
Finding #4: Four Patterns Separate Winners Regardless of AI Score
Across all 35 SERPs, we found four structural patterns that showed up in 82 percent of top-three-ranking URLs, regardless of whether the page itself contained AI-assisted copy.
First, named authors with verifiable credentials. Brian Dean at Backlinko, HBR writers with published books, founder-operators who document their company role, and senior practitioners who state their professional experience. A named author with a bio that can be checked is worth more than a clean-AI byline. Second, proprietary data. Original surveys like Siege's 353-marketer study, Paddle's VC data, Mixpanel's 12,000-company sample, and Salesforce's FY25 metrics from 202 SMBs all outrank secondary analysis. Google seems to treat "data that cannot be synthesized from other public sources" as a strong ranking signal. Third, named customer examples with specific numbers. HubX recovered "$106,000 in 3 months, 63 percent retention." Salesforce's case studies name Checkwriters with "20 percent revenue growth" and Liquidity Services with "50 percent cost cut." Generic "our customer saw a 40 percent increase" language is a middle-of-page-one pattern. Named customer plus specific number is a top-three pattern. Fourth, first-person operator voice with quantified proof. "I've used this since 2016, sent 640,199 emails." "Led SEO at Webflow, a $4.2 billion company." "350 hours of personal testing." "$40k from SEO in 2025." This is the pattern that AI content cannot fake, and it is the pattern that dominates commercial top-threes.
Teams building for ranking should treat these four patterns as the scoreable checklist, not the AI score. To understand how this translates into specific on-page optimization for AI search engines as well, our breakdown of GEO and SEO working together goes deeper into the operator-voice and named-data pattern.
Finding #5: Seven AI Tells That Show Up in Almost Every Low-Ranking Page
We tracked tells across all 350 URLs. The following seven appeared in 89 percent of pages scoring 16 or higher.
Triplet verb constructions
"Automate, personalize, and measure." "Streamline, optimize, and scale." "Attract, engage, and convert." The three-verb pattern is so consistent across AI-written copy that its presence alone is a moderate AI signal.
Empower, leverage, transform, elevate, unlock, navigate
The vocabulary cluster of AI marketing-speak. A page that uses three or more of these verbs in a 2,000-word article is almost always AI-assisted at minimum.
"In today's [X] world" or "in today's fast-paced digital landscape"
The universal AI opener. It signals the page did not start with a specific observation or a specific reader question.
"Not a one-size-fits-all"
Usually followed by a listicle that in fact offers a generic one-size-fits-all framework. The phrase is a hedge, not a differentiator.
Round-number perfect-count enumeration
Exactly 7, exactly 10, exactly 12 items with no rationale for the count. Human-edited content often has 6, 8, 11, or 14 items because that is how the argument actually shaped up. AI drafts default to round numbers.
Recursive definition padding
Restating the query as a question inside the answer. "What is marketing automation? Marketing automation is..." or "How do you improve conversion rates? Improving conversion rates starts with..."
Generic case study language
"One of our customers saw a 40 percent lift." Without a customer name, timeframe, or specific context, the claim is either unverifiable or fabricated, and Google appears to notice. The switch from generic case study language to named, numbered case study is one of the biggest delta-makers between mid-page-one and top-three ranking.
Finding #6: The Rank Differentiation Between Top-3 and Positions 4-10
Two of the most interesting corpus-wide numbers came from comparing the top three rankings to positions 4 through 10 within the same SERPs. Top three URLs averaged an AI score of 5.4. Positions 4 through 10 averaged 8.1. Top three URLs contained fully AI-generated content (score 16 or higher) about 4 percent of the time. Positions 4 through 10 contained that same level of AI content about 18 percent of the time.
Google's ranking algorithm is clearly differentiating within each SERP. The top spots reward something that looks more human-verified, while mid-page slots are more AI-tolerant. For SEO teams, this means that "good enough" AI content can get to page one, but cannot usually break into the top three. The jump from position 6 to position 2 is not an incremental improvement in AI-polish. It is a step change in operator voice, named data, and verifiable credentials.
The implication for algorithm-resilient SEO is straightforward. Content optimized for the middle of page one will never stop being AI-assistable. Content optimized for the top three needs to survive a different test, which is whether a human reader can tell the writer is real, the data is proprietary, and the examples are verifiable.
What This Means for Your Content Strategy
The practical takeaway from 350 scored URLs is not that AI content is failing or winning. It is that AI tolerance is query-specific, and the four winning patterns apply everywhere. If you're running a commercial "best X" page, the top three slots are almost impossible to hit without a named operator reviewer and verifiable proof. If you're running a definitional page, AI-assisted copy is fine as long as the domain carries weight and the content is accurate. If you're running a technical-definitional page for developers, write it like a developer, cut the marketing language, and cite sources. If you're running a trend or listicle page, invest in proprietary research, because a 353-person survey will out-rank a hundred links every time. If you're running professional services content, put a named author with real credentials on it.
The teams that will lose over the next 18 months are the ones running an AI-content playbook for every query type. The teams that will win are the ones who know which slot on which SERP tolerates which kind of content, and who invest their editorial effort in the four patterns that actually differentiate winners. Our work with clients has driven 750 percent average AI visibility growth by applying exactly this kind of query-category thinking across both Google and the AI engines, because the winning patterns are increasingly the same in both.
Ready to Stop Guessing What Google Actually Rewards?
If you want to apply the patterns from this 35-SERP study to your own content before your competitors do, our team can audit your current rankings against the four winning patterns and show you exactly where your pages sit on the AI-signal spectrum relative to the top three in your category. Book a strategy call with Passionfruit and we will walk you through a query-category gap analysis for your top revenue-driving keywords.
Frequently Asked Questions
Does Google penalize AI-generated content in 2026?
Not categorically. Our 35-SERP study found AI-generated content ranking in the top 10 for almost every category, with fully AI-generated pages holding positions 4 through 8 in roughly 18 percent of SERPs analyzed. Google's helpful content guidance focuses on whether content is useful to readers, not whether it was AI-assisted. The correlation between AI-signal score and rank position in our data was weak at around -0.18, meaning AI signal alone does not predict ranking outcomes.
What is the single biggest differentiator between top-three and mid-page content?
Named authors, proprietary data, and named customer examples with specific numbers. These three patterns appeared in 82 percent of top-three URLs and only 31 percent of positions 4 through 10. First-person operator voice with quantified proof ("I've sent 640,199 emails," "led SEO at a $4.2 billion company," "$40k from SEO in 2025") was the single strongest pattern in commercial "best X" SERPs.
Can AI-assisted content still rank in the top three?
Yes, but only for specific query types. Definitional queries and enterprise comparisons tolerate AI-assisted copy in top ranks when the domain itself carries authority. For commercial "best X" queries, professional services queries, and technical-definitional developer queries, AI-assisted content struggles to break the top three without the four winning patterns layered on top.
How do I tell if a page is AI-generated without a detection tool?
Look for the seven tells: triplet verb constructions, the empower-leverage-transform-elevate-unlock-navigate vocabulary cluster, "in today's X" openers, "not a one-size-fits-all" hedges, round-number perfect-count lists, recursive definition padding, and generic case-study language with no named customers or specific numbers. Three or more of these in a 2,000-word article is a reliable indicator of heavy AI assistance.
What about AI search engines like ChatGPT and Perplexity? Do the same patterns apply?
Largely yes, with some amplification. AI engines appear to reward proprietary data and named authority even more heavily than Google does, because citation is a first-class concept in AI answers. Our schema markup guide for AI search covers how structured data affects citation frequency in AI engines, which matters more than traditional rank for many buyer journeys.
How can I replicate this study for my own niche?
Pick 10 to 15 queries across the intent categories relevant to your business, fetch the top 10 URLs for each, and score every page on the eight AI-content dimensions we used. Compare your current ranking content to the four winning patterns. The gap between where you score and where the top-three in your category score is your optimization roadmap.
