What have the authority sites measured recently?

On 2026-07-06, Ahrefs published “Google Doesn’t Punish AI Content; It Punishes Bad Content.” They compared 331,000 pages to test a widely repeated claim: that using AI generation makes rankings fall. They found no evidence of pages being acted on “because they were AI-generated”; what got acted on were pages that were low quality and unhelpful to readers.

This conclusion actually matches what search engines have said publicly for years. A search engine does not care what tool you write with; it cares what problem the page solves, whether it holds unique information, and whether it is fishing for clicks. But many people now quote only the first half, turning it into “AI content is safe, so you can roll it out at scale.” That reads a research finding as a get-out-of-jail-free card.

I think the more useful reference is another content experiment the same firm ran in 2026. They compared “self-promotional” and “useful” approaches side by side, and the results split again. Read together, the two pieces point to one judgment: the writing tool does not matter; whether you give the reader something extra while writing does. I also develop this thread in the relationship between content and GEO.

What is this conclusion most easily misread as?

The most dangerous misreading is this: since AI content is not punished, open ten accounts and generate fifty posts a day. That approach already failed once, in the keyword-stuffing era a few years ago. What actually happens is that your site gains hundreds of near-identical pages with repeated information, drowning the few pages that are genuinely useful. We wrote about exactly this problem in publishing more posts does not equal getting more traffic.

The second misreading is “AI writes better than people do, so we can hand it everything.” My own testing gives me the opposite impression. AI is good at organizing known information into smooth prose; it is bad at supplying information nobody has. Ask it to write about “common defects in hardware surface treatment” and it produces a textbook paragraph. Ask it to write “last year this batch of plating at our plant blistered, and we traced it to a pickling step in pre-treatment that ran two minutes short,” and it cannot, because it does not know. That specific experience is the most valuable part of your page.

So I state AI’s role plainly: it is a wordsmith, not a source of information. The person supplying the material has to be you, and the person judging what is right has to be you too. If you do not say at the end of an article where the information came from, both readers and search systems can tell.

What counts as bad content? The three most common kinds on export sites

These three kinds share one trait: swap out the company name and the page still stands. Any passage you can replace wholesale with a competitor’s name is basically bad content. Conversely, as soon as a number, a condition, or a point in time appears that only your company knows, the page has a chance of being cited and remembered. You can use how to judge thin content and soft 404s to health-check it page by page.

What do our own indexation measurements show?

In 2026 I ran an indexation observation on a set of canary pages. The seven pages took until day 14 to reach 7/7 indexed, but only 2/7 ever appeared in impressions. At the same time, this site’s overall indexation rate was 100%, yet it drew only 12 impressions and 0 clicks in 28 days. Put those two numbers side by side and they say plainly how much “getting indexed” is really worth.

The old anxiety in export circles was “the site is not getting indexed,” so some people built site networks and scraped content. Indexation is basically no longer the bottleneck; the bottleneck is what comes after: does anyone search, will they click once they find you, and does a click turn into an inquiry. If you still use “number of pages indexed” as your KPI, you will feel you are doing well; switch to impressions, clicks, and citations and the picture goes cold at once. I keep the related method notes in field research and method notes.

This also explains why mass generation makes less and less sense. The pages you work hard to generate may never earn a single impression; all they contribute is an indexation number. Spend that effort instead on ten pages of in-depth content carrying real experience, and the result is usually better. See content pruning and cleanup: cut the weak pages first, then talk about adding new ones.

How should AI actually be used?

Use it for structure and first drafts, not for facts

I first write a one-page outline: the question to be answered, the key facts I know, the exact words clients keep asking. Then I have AI expand it into smooth paragraphs. The facts have to come from me; the expansion can be delegated. A draft produced this way carries information only we have, rather than a generic template. For the writing method, compare how to write an answer passage that can be cited.

Use it for cleanup, not for volume

The more cost-effective use is the reverse: hand old pages to AI and have it point out where they are empty, where they repeat, and where readers will get lost. Cleaning up old content often pays better than adding new content. Combine it with content updates and refreshes, and start by rescuing the pages that already earn impressions.

Every page needs a named person accountable for it

No matter who wrote it, the page must show a real person accountable for the information. A byline, an update date, and a contact entry point — these three make responsibility concrete and also affect citation judgments. The cheaper the tool, the more expensive the responsibility.

A low-cost action checklist

Three steps, no cost, doable this week.

1. Pick 10 old pages for a health check: ask of each one, “Does it still stand if I swap out the company name?” Mark the ones that do as needing revision.
2. Add one piece of unique information to each of the 10 pages: a real number, a failure case, a delivery record — any one will do.
3. Build a fact-supply sheet: collect the information only you know. From now on, start every draft from that sheet and let AI handle only the wording.

Following the direction of Ahrefs’ 331,000-page study, quality is the only stable variable. To push ahead systematically, see how page structure affects the probability of being cited and put your unique information where it is easiest to lift out.

The questions owners ask most often

Will Google punish export articles written with AI?

According to Ahrefs’ 2026-07-06 study of 331,000 pages, what gets flagged is content quality, not the method of generation. AI is only a tool; writing something empty is the real risk.

If AI content is not punished, can we mass-generate hundreds of pages?

Not advisable. Mass generation most easily produces thin pages that repeat and add nothing new, and those are exactly what gets acted on. They also dilute the content on your site that is genuinely useful.

How do you tell whether a page is thin content?

Ask three questions: does this page contain specific information found nowhere else? Does it give the conditions and numbers where it applies? If you deleted it, would it affect a buyer’s decision? If the answer to all three is no, it is basically thin content.

Our industry has little published material. What if we cannot write anything in depth?

Depth does not have to mean length. It can be your own production-line details, rework records, or how a quote is built. Others cannot copy those, and AI cannot invent them.