AI-Generated Content and the Quality Collapse Ahead

8 min read · 1,658 words

The quality problem with AI generated content is not a prompting problem. It is an economics problem. When publishing costs almost nothing, volume expands until attention, not production, becomes the binding constraint. More than half of new web articles are now machine-written, and almost none of them get read.

I run a content operation, and AI is in it. Drafting, summarising, first-pass desk research, translating copy for markets whose language I don’t speak. So take what follows as a note from inside the machine rather than a purity argument from someone who has never used the tools.

The habit I’m arguing against is older than the tools. For eight years I ran marketing across Asia, the Middle East, Africa and Turkey, and in four of those markets I owned a content calendar that was, honestly, a treadmill. Roughly a third of what went out earned any measurable reader response. The rest existed because the calendar said something was due on Thursday. AI didn’t invent that problem. It removed the only thing that had been limiting it, which was the cost of writing the Thursday post.

The cost of publishing fell to zero, and volume did what volume does

Graphite analysed 43,000 randomly sampled English-language articles from Common Crawl and found that AI-generated articles overtook human-written ones on the open web in November 2024, and have hovered near half of all new output ever since. Ahrefs scanned 900,000 newly created pages in April 2025 and found 74.2% carried at least some machine-written text, with only 25.8% purely human. The Content Marketing Institute’s survey of 1,015 B2B marketers puts organisational AI use at 95%, while only 59% rate their content marketing as even somewhat effective.

Read those together and the story is plain. Adoption is total, output has roughly doubled, and self-reported effectiveness has not moved with it. That is exactly what an economist would predict. Once the marginal cost of producing one more article approaches zero, a rational team produces articles until some other constraint bites. The constraint that bites is not budget or headcount any more. It is the number of minutes a buyer will give you, and that number has not grown since 2019.

Most of the industry has diagnosed this as a craft failure and prescribed better prompting, better editing, better brand-voice files. I’ve written those files myself. They improve the artefact. They do nothing about the arithmetic, because the arithmetic is about everyone else’s output, not yours. My view of what AI has genuinely changed in brand work is that it compressed execution cost and left judgement exactly where it was.

Does Google actually penalise AI-generated content?

No, and marketers keep misreading that as good news. Ahrefs examined 331,000 pages across 100,000 search results in June 2026 and found 5.3% of positions one to three were entirely machine-written, while 82.2% of top-three results contained under 50% AI text. Pages with light AI use drew two to three times the organic impressions of heavily generated ones, and indexation rates slid from 49.28% for low-AI pages to 40.35% for very high ones.

Google’s spam policies define scaled content abuse as generating many pages “for the primary purpose of manipulating search rankings and not helping users”, and list generative AI as one method among several. The tool isn’t the offence. The scale-without-value is.

So there is no penalty for using AI. There is also no reward. If a machine can produce your article, a machine can produce a hundred competing versions of it by Friday, and none of you gains position. Absence of punishment is not the same as advantage, and a strategy built on the first is not a strategy.

Where does all that content actually go?

Mostly nowhere. Graphite’s researchers noted that the AI-written articles they identified largely fail to surface in Google or ChatGPT at all. And the exit door is narrowing for everybody, human-written work included. SparkToro’s analysis of Similarweb panel data found 68.01% of Google searches ended without a click in the first four months of 2026, against 60.45% in 2024.

Pew Research Center tracked 68,879 searches made by 900 US adults and found users clicked a traditional result 8% of the time when an AI summary appeared, against 15% when it didn’t. Clicks on a source inside the summary happened in 1% of visits. Sessions ended outright 26% of the time on pages with a summary, versus 16% without. Ahrefs measured a 58% fall in click-through at position one when an AI Overview is present, up from 34.5% eight months earlier.

The picture is a production line running at double speed into a warehouse whose doors are closing. Publishing more into that is not ambition. It is a way of losing money quietly, in instalments small enough that nobody calls a meeting about it.

Attention is the only input that did not get cheaper

Every other input to content marketing has deflated. Writing, design, translation, research summarisation, video editing. Attention has not. It is the one genuinely scarce factor left, and scarce factors capture the value in any system where everything else is abundant. That single sentence should be reorganising content budgets, and in most companies it isn’t.

Spend keeps climbing anyway. Dentsu’s India digital advertising report puts Indian digital ad spend at Rs 71,621 crore in 2025, up 19% year on year, heading for Rs 98,034 crore by 2027 and close to 70% of total advertising. More money chasing fewer clicks, with production costs falling. Something in that equation has to give, and it will be the assumption that output correlates with outcome.

The Indian brands that already understand this are the ones publishing least. Zerodha’s Varsity is a small library of genuinely instructional material that has outlasted a decade of blog calendars. Asian Paints built Beautiful Homes as a destination rather than a feed. Zoho’s documentation is more useful to a buyer than most of the marketing built on top of it. None of them wins on volume. They win because a person who reads one piece has a reason to come back for the next.

What do AI engines actually cite?

They cite things mass production cannot fake. Ahrefs studied 75,000 brands and found branded web mentions correlated with AI Overview visibility at 0.664, roughly three times the correlation of backlinks at 0.218. Brands in the top quartile for web mentions averaged 169 AI Overview mentions against 14 for the next quartile down. Pew found 88% of AI summaries cite three or more sources, with a median summary length of 67 words.

Underneath that sits Google’s own helpful content guidance, which asks whether a page provides “original information, reporting, research, or analysis” and whether it demonstrates “first-hand expertise” from having actually used the product or visited the place. It also asks who wrote it, how it was made, and why it exists. Those are precisely the four things an unattended generation pipeline cannot supply: a named person, a method, a reason, and a fact nobody else has.

This is the mechanism people miss when they talk about getting cited by AI engines. The filters are not aesthetic. They are provenance filters, and provenance is the one attribute that gets more expensive as content gets cheaper.

Production budget versus research budget

Here is the same money spent two ways. I’ve run both models, and the second one is harder to defend in a quarterly review and easier to defend in a two-year one.

The volume column is not worthless. It has a job, which is coverage of the questions buyers ask on the way in. The mistake is running the whole function that way and calling the output a brand.

What I would do with next year’s content budget

Five moves, in the order I’d make them.

1. Cut published output by half and hold the budget flat. The saving is not a saving, it is a transfer. Every rupee released from production goes to research.
2. Buy one piece of primary data a quarter. A survey of 300 customers in your category, a teardown of pricing across your competitive set, an analysis of your own service data that nobody outside has seen. One original number is worth thirty explainer posts, because thirty explainer posts already exist.
3. Put a named human on every piece. Author bylines, a real biography, a track record that resolves to a person. Both Google’s guidance and the citation data reward attributable expertise, and a pipeline cannot manufacture a reputation.
4. Use AI where it is genuinely good. Structuring, summarising, translating, first drafts of things that are not the argument, and analysis of data you already own. Keep it away from the thinking and the claim.
5. Change the metric. Stop reporting pieces published. Report citations earned, mentions won, and named accounts that engaged. If you are rebuilding the budget case from scratch, my working view of what a B2B company should actually spend on brand is the place I’d start the conversation with finance.

None of this needs an AI policy document. It needs one decision: that the content function is measured by what it discovers, not by what it emits.

My honest position is that the coming shakeout will be good for anyone who does original work and brutal for anyone whose content operation was always, underneath, a volume operation with a nicer name. The tools didn’t cause that. They just removed the friction that was hiding it. If you cut your publishing schedule in half tomorrow and spent the difference on finding out something true about your market, would anyone who reads your content actually notice the drop?