Google has never said it penalises content for being written by AI. It has said the opposite, in writing, repeatedly. And yet almost every programme built to avoid an "AI penalty" is built to avoid the wrong thing, which is why so many of them fail anyway.

The sentence that matters is in Google's spam policies. The old prohibition on "spammy automatically-generated content" was removed in March 2024 and replaced with a scaled content abuse policy covering pages generated primarily to manipulate rankings rather than help users. Google's own announcement states the policy applies no matter whether content is produced through automation, human effort, or some combination of the two.

Method is irrelevant. Purpose and scale are what get assessed. That single distinction changes what you should build, and it is where our view of search in 2026 starts.

Is there actually any evidence AI content ranks worse?

No, and the best available data says the ranking web is already hybrid.

Ahrefs ran 100,000 random keywords, pulled the top 20 results for each, and put roughly 600,000 pages through its own AI-detection model in July 2025. The results:

• 4.6% of pages were classified as purely AI
• 13.5% were classified as purely human
• 81.9% were mixed — some AI involvement, some human
• 86.5% of top-ranking pages contained some AI-generated content
• Correlation between AI content share and ranking position: 0.011, effectively zero

Read the honest version of that: almost nothing ranking is purely machine-written, almost nothing is purely hand-written, and the proportion of AI involvement does not predict position. Note the limitations Ahrefs states itself — AI detectors are probabilistic and carry false-positive risk, and Ahrefs sells SEO software. Treat it as the strongest available evidence rather than proof.

So what actually got hit in 2024?

Scale, not AI. But the correlation was near-total, and that is exactly why the folklore took hold.

Practitioner trackers monitoring the March 2024 updates found that essentially all the sites they observed being deindexed or receiving manual actions contained AI-generated posts, with many consisting almost entirely of them. The industry read causation into that. The simpler explanation is economic: every operation mass-producing junk pages in 2023 and 2024 used AI because it was the cheapest way to reach volume. Google targeted the volume and the absence of value. The AI was incidental to the spammer's cost structure.

Two mechanisms get merged here and they are not the same thing, which matters because the remedies differ:

Algorithmic demotion. A core or spam update reassesses quality and your rankings fall. You receive no notification. There is no form to submit. Recovery, where it happens, comes from the next reassessment after the underlying content changes — and plenty of sites hit in 2023 and 2024 never recovered at all.

A manual action. A human reviewer takes action and you get a notice in Search Console, with a reconsideration process. This is rarer and it is the only case where "penalty" is the right word.

If your traffic fell and Search Console shows no manual action, you were not penalised. You were reassessed. Anyone selling you a penalty recovery service without checking which happened is selling you the wrong service.

What is the Unrepeatable Element?

Here is the editorial gate we apply, and it is one question per page: can you name the single thing on this page that a model could not have generated on its own?

Original data you collected. A process your team actually ran and can describe. A number from your own operations. A named expert's judgement on a specific question. A photograph of the real thing. A customer problem you have handled forty times and can describe precisely.

If you cannot name it, the page is a paraphrase of what already ranks, and it is competing against a system whose entire purpose is to summarise what already ranks. Publishing it adds a page to the scaled-content side of the ledger and buys you nothing.

Is that Google's position or is it industry folklore? Partly both, and it is worth being exact because the distinction is widely fudged. Google's guidance on creating helpful, reliable, people-first content asks self-assessment questions about whether content offers original information, reporting or analysis, whether it provides insight beyond the obvious, and whether it merely summarises what others have said. Its quality rater guidelines direct raters to assess the extent to which content offers unique, original information or perspectives not already available elsewhere. So the principle is documented in Google's quality material.

What is not documented is any mechanism by which a system detects first-hand experience. It is described as a quality to aim for, not a signal that is measured. An honest article says that rather than implying Google has an experience detector.

Did Google just tell everyone to stop doing GEO?

Broadly, yes — and this is the most useful thing published on the subject this year.

In May 2026 Google released a resource on optimising for generative AI features in Search, announced by John Mueller. Its position is that these features are rooted in Google's core Search ranking and quality systems, and that existing SEO practice still applies. Semrush's summary lists the tactics Google says can be ignored:

• AI-specific rewriting for every long-tail phrasing, because the systems understand synonyms and meaning
• Content chunking — splitting articles into fragments — because the systems can extract relevant passages from multi-topic pages
• llms.txt files, which Google's crawler treats like any other text file

If an agency has quoted you for an "AEO layer" or a "GEO retainer" on top of SEO, that quote is for three of the things Google has just said to ignore. Search visibility is one discipline, not three. We have been saying that in our work on AI-powered marketing, and it is now Google's published position rather than our opinion.

But do AI answers mean content gets no clicks anyway?

This is the serious objection, and it deserves the numbers rather than a reassurance.

Pew Research Center tracked the actual browsing of 900 US adults across roughly 69,000 searches in March 2025. The findings:

• Visits where an AI summary appeared included a click on a traditional result 8% of the time, against 15% where no summary appeared
• Clicks on links inside the AI summary: about 1% of visits
• Sessions ended entirely on 26% of visits with an AI summary, against 16% without

Google disputes the study, calling the methodology flawed and the query set unrepresentative of Search traffic. We think the study is the best independent measurement available and that Google's objection is worth stating rather than hiding.

Then the headline number everyone quotes, with the caveat nobody quotes. SparkToro's clickstream analysis puts US zero-click searches above 68%, up from around 60% two years earlier. But that definition counts any search not ending in a click to the open web — so a trip to Maps, a YouTube video, or simply refining the query all count as zero-click. Semrush's stricter taxonomy, which separates query refinements, yields a far lower desktop figure in the mid-twenties. Both are correct measurements of different things. For a business that only earns from traffic to its own domain, the broader definition is the one that describes your reality.

The redistribution finding is the one that should shape your plan. Analysis by Fractl across more than a million high-volume keywords found informational queries declining substantially while total search demand stayed roughly flat. Search did not shrink. It moved. Top-of-funnel "what is" and "how does" demand migrated to assistants; complex, comparative and lower-funnel querying held up or grew.

And penetration is wildly uneven by intent. AI summaries appear on the large majority of commercial-research queries — the "best X", "X vs Y", "reviews" phrasings — while direct ecommerce product searches trigger them on a small single-digit percentage of queries, and local service queries only slightly more. Which gives you an unusually clear instruction: the further down the funnel and the more local your query set, the less of this applies to you. It is also why local search has quietly become the most defensible organic asset a service business owns.

What does the editorial layer actually consist of?

"Human review" is the phrase every agency uses and almost nobody defines. If the throughput gain from generative tools is real, the recovered hours have to go somewhere, and this is where. Four steps that are process rather than principle:

Claim verification before publication. Every factual assertion, figure and citation checked against a primary source by someone who is accountable for it. Models fabricate citations confidently and the failure is invisible to a casual reader, so this is not optional on anything with a number in it.

The element, named and logged. Before a piece is scheduled, the person publishing it writes down what the unrepeatable element is. If the field is blank, the piece does not publish. This is the only control that reliably stops volume drifting into the scaled-content category, because it fails closed.

Named human attribution. A real person with a real role, who could defend the content if asked. Google's 2023 guidance advised against crediting an AI model as an author, and the practical reason stands on its own: an unattributable page is one nobody is answerable for.

A review cadence with ownership. Each cluster has an owner and a revisit date. Content ages, regulation moves, and a page that was accurate in March can be wrong by October — a problem this series has run into repeatedly, since several of these articles describe rules that changed within the year.

None of that requires software. It requires someone's name against each step, which is why it is the part most programmes quietly skip. The services we run on this are deliberately boring for the same reason.

How do you triage an archive you have already published?

You have four hundred pages and you cannot delete them on a hunch. Three options per page, chosen page by page rather than site-wide:

Add the unrepeatable element. Best option for a page targeting a query you genuinely want. Put in the original number, the named expert, the process description, the photograph. This is work, and it is the only option that improves the asset.

Consolidate. Several thin pages circling one topic become one page that has the element. Redirect the rest. This is usually the highest-leverage move on an AI-produced archive, because such archives are typically wide and shallow by construction.

Remove. For pages that are obsolete or serve no user. Note Google's own caution here: removing pages because traffic is low is not the same as removing pages because they are worthless. A page serving a narrow, low-volume intent may be doing its job. Traffic thresholds are a bad guillotine, and bulk deletion driven by a dashboard is how businesses destroy working long-tail visibility.

Practical sequence we use: group the archive by intent cluster, find the clusters where you have something unrepeatable to add, fix those first, consolidate the thin remainder within each cluster, and only remove what is genuinely dead. Expect reassessment to take time and to arrive without announcement — which is why measuring this properly means tracking cluster-level impressions and conversions rather than watching a single page.

One governance point worth adding. Keep a record of how each piece was produced, who reviewed it, and what the original element was. That record is useful for quality control, and it is increasingly useful for other reasons too, since disclosure expectations around AI-produced marketing material are tightening and the analytics and documentation layer is where it should live.

Two things that will get you hit that have nothing to do with AI

Site reputation abuse. Enforced from May 2024, this targets authoritative sites hosting low-value third-party pages that free-ride on the host's signals. If you run a partner section, a syndicated feed, a guest-post programme or a coupon subdomain without real editorial control, that is your exposure, and it is unrelated to whether a model wrote anything.

Publishing at scale against query permutations. The classic programmatic build — one template, a spreadsheet of city and service combinations, a thousand pages. This is the behaviour the scaled content abuse policy describes, and doing it by hand does not help. Google's 2023 statement on AI and content framed the test as whether AI is being used as a cheap shortcut to game rankings or as a tool to help produce something useful for people. That framing has not changed in three years.

When you should not build a content engine at all

Three situations where we would tell you to spend the money elsewhere, and we would rather say it now than in month four.

You have no original element available and no path to one. If your business holds no proprietary data, employs no recognised expert, and cannot describe a process others cannot, then "add original insight" is a counsel of perfection rather than advice. For a reseller or a small service business, the honest recommendation is to skip informational content and put the budget into demand capture — local search, performance marketing, and the lower-funnel pages where AI summaries barely appear. That is a real strategic answer, not a failure.

Your query set is informational and top-of-funnel. That is precisely the demand migrating to assistants. Measuring click-through on those queries before committing a year of budget is cheap, and it is the test most agencies skip because they already know the answer.

Your binding constraint is the offer or the sales process. No quantity of excellent content fixes a weak proposition, a slow follow-up or an unclear price. Content makes a working funnel bigger and a broken one more expensive, which is the mechanism behind most lead generation programmes that fail.

A word on the broader context, because it bears on how long any of this holds. Penske Media filed an antitrust suit against Google over AI summaries in September 2025, and the European Publishers Council filed a formal complaint with the European Commission in February 2026. Those actions may eventually change what publishers can opt out of without losing search visibility. They are worth watching, and they are not a reason to delay a decision now.

Want your archive triaged rather than guessed at? Send us your site and your Search Console access and we will come back with the clusters worth fixing, the pages worth consolidating, whether you have a manual action or an algorithmic reassessment, and an honest answer on whether your query set is one that organic content can still win. If it is not, we will say so and recommend where the budget should go instead. Talk to Midgrow, or message +91 9632398896. What this work costs is set out in what agencies do not tell you, and how engagements are structured is on our digital marketing services page.

Frequently asked questions

Does Google penalise AI-generated content?

No, not for being AI-generated. Google replaced its automatically-generated content policy in March 2024 with a scaled content abuse policy covering pages produced primarily to manipulate rankings rather than help users, and the policy explicitly applies whether content was made by automation, humans, or a combination. Purpose and scale are assessed, not production method.

Can AI-written content rank on Google?

Yes, and most ranking content already involves some. An Ahrefs study of around 600,000 pages across 100,000 keywords found 86.5% of top-ranking pages contained some AI-generated content, only 4.6% were purely AI and 13.5% purely human, with a correlation between AI share and ranking position of 0.011 — effectively zero.

What is scaled content abuse?

Generating many pages for the primary purpose of manipulating search rankings rather than helping users. It is the policy that replaced Google's older rule against automatically generated content. Programmatic page builds against keyword permutations are the classic example, and producing them by hand rather than with AI does not change the assessment.

How do I know if I was penalised or just lost rankings?

Check Search Console for a manual action notice. If there is one, you have a penalty and a reconsideration process. If there is none, your rankings were reassessed by a core or spam update — no notification, no appeal, and recovery depends on the next reassessment after the content changes. The remedies are different, so establish which happened first.

Do I need a separate AEO or GEO strategy for AI search?

Google's May 2026 guidance says its generative AI features are rooted in its core Search ranking and quality systems and that existing SEO practice applies. It specifically lists AI-specific rewriting, content chunking and llms.txt files among tactics that can be ignored. A separate retainer line for those is a charge for things Google has said do not help.

How much traffic do AI Overviews actually take?

Pew Research Center's tracking of 900 US adults across about 69,000 searches in March 2025 found clicks on traditional results fell to 8% when an AI summary appeared, against 15% when none did, with 1% clicking links inside the summary and 26% of such sessions ending entirely. Google disputes the methodology. Penetration varies sharply by intent, with commercial-research queries heavily affected and ecommerce product and local service queries far less so.

Should I delete old AI-generated pages?

Not on a traffic threshold. Google cautions against removing pages simply because traffic is low, since a page serving a narrow intent may still be useful. Work cluster by cluster: add an original element where the query matters to you, consolidate thin pages into one stronger page, and remove only what is genuinely obsolete.

Is text content covered by India's AI labelling rules?

India's notified definition of synthetically generated information covers audio, visual and audio-visual content, so AI-assisted written copy sits outside those specific labelling duties. Advertising standards still apply to what the text claims, and keeping a production record is sensible practice regardless.