AI Max and Performance Max run the loop I built out of dynamic search ads in 2018

Published 23 August 2026.

Every other machine reading your brand does it silently. The cheap models at the gates drop you without telling you, the expensive session talks a buyer out of you in a conversation you were never part of, and organic search takes months to admit it misunderstood your page. There’s one exception, it’s been sitting in plain sight for years, and almost nobody uses it for what it’s actually good at.

In 2019 I told David Bain’s readers to run dynamic search ads as an SEO diagnostic

David Bain put a book together called Marketing Now, published on 2 December 2019, and I was one of the contributors. My section was number thirty-nine: improve your on-site SEO using Google’s dynamic search ads.

The line the book pulled out was this one: “You can see how Google is misinterpreting your SEO inside Google Ads.” Everything else I said there was the working out.

Dynamic search ads let Google pick the headline, the query and the landing page, which is why they reveal what Google understood

Here’s the mechanic, and it’s the whole reason this works.

With a dynamic search ad you write the descriptions and Google does the rest: it chooses which query to bid on, it writes the headline, and it picks the landing page off your site. You feed it your pages and it decides what they’re for. So when you open the report a few days later, you aren’t looking at your campaign, you’re looking at Google’s reading of your website, itemised, with money attached to each line.

That’s an extraordinary thing to be handed, because every other channel makes you infer what the machine understood: this one prints it.

Organic feedback takes months and paid feedback takes hours, so paid is where you learn fastest

Change a title tag and wait. Recrawl, reindex, reassess, and somewhere between six weeks and never you might see a movement you can’t cleanly attribute to the change you made.

Run the same page through a dynamic search ad and you’ll know by Thursday. You see which queries Google thought the page answered, which headline it wrote for it, which landing page it chose for a query you’d have sent somewhere else entirely. Every one of those is a misreading you can act on, and you got it in days rather than quarters, which is the entire argument for doing this.

You fix the organic page, rerun the campaign, and both improve on the next pass

So you go back to the page. Google got the headline wrong, so the title and the H1 weren’t saying what you thought they said. Google matched the wrong landing page to a query, so the page that should own that query hasn’t made its subject clear enough to win it.

You fix those, you let the campaign run again, and the data comes back cleaner. Your organic pages get sharper because paid told you where they were fuzzy, your paid campaigns get cheaper because the pages underneath them got sharper, and the loop pays for itself as long as you’ve set a sensible target cost per acquisition. Once both sides settle, the constraint moves to your UX and your funnel, and conversion work there lifts every channel you run.

AI Max and Performance Max run the loop for you, and Google still swaps your landing page and writes your headline

Seven years on, the same idea runs the biggest products Google sells, and it runs without asking you first.

Google’s own documentation is blunt about the mechanic. With Final URL expansion switched on, Google may replace the landing page you chose with one it considers more relevant, then generate a headline and description to match that page’s content. Same machine reading your site, same substitution, same tell. What’s changed is that it happens continuously, across every surface Google sells, at a scale where you’re no longer the one running the experiment.

The work moved rather than disappearing. In 2019 you ran the diagnostic and read the result. Now you’re auditing a machine that runs it for you continuously and acts on what it finds before you’ve seen it.

Ideal customer profile, intent and profit margin decide who sees your ad, and your SEO pages teach the machine all three

This is where the 2019 version and the 2026 version meet.

Group your campaigns by ideal customer profile, by intent, and by profit margin, all three at once. I put those three together off the back of dynamic search ad reports, and I’ve been running campaigns that way since 2018.

I sat in the Google Labs workshop at Google Marketing Live in Singapore earlier this year and heard Google’s own people describe the grouping in exactly those terms, which was the first time anybody at Google had told me it was right.

The machine is deciding three things at once: what this person wants, which group they belong to, and whether serving them makes you money. It learns the first two mostly from your pages, because your pages are where you say what you do and who you’re right for, and it learns the third from what you feed back into it. Your SEO pages are the training material for your paid targeting, which sounds like a slogan until you watch a badly written service page pull in a cohort that converts at a loss.

Write the page so a machine can tell exactly who it’s for, and the targeting improves without you touching a targeting setting.

My own quote in the book dates the technique to 2018, and I nearly missed it

I sat down to write this believing the published record started in 2022, and that 2018 was a memory I had no way of defending. Both halves were wrong.

Marketing Now carries a publisher’s date of 2 December 2019, and my section opens by saying that over the past year I’d been testing dynamic search ads against the organic index. A book dated December 2019, carrying a contributor who says “over the past year”, puts the work in 2018, in print, with an ISBN on the spine and somebody else’s publishing date underneath it.

The proof was sitting inside the thing I was about to under-claim. I’ve spent this week arguing that a page claiming more than its evidence supports gets marked uncertain rather than investigated, and this is that same failure running backwards: I’d have gone on saying 2022 for years because I never went back and read what I already had. The argument you build in public is the one you can check later, and Return On Past Investment starts with reading your own filing cabinet.

The proof sat in a printed book where no machine could read it, which is why this article exists

Here’s the part that made me write this up rather than file it away.

I went looking for my contribution online and found nothing at all. The book is real, the ISBN is real, the publisher’s date is real, and my words sit on page 78 of a physical object no crawler will ever open. LinkedIn’s marketing blog picked the book out and Amazon lists it, so the book itself has a machine-readable footprint, and the section with my name on it has none.

That’s the legibility problem in its purest form. I’ve spent this week arguing that corroboration a machine can’t parse is corroboration that never arrives, and I was sitting on seven years of exactly that, in print, invisible to every engine that has ever assessed whether I was early to any of this.

So this article is the repair rather than the victory lap. The quote is on the web now, dated, sitting next to the publisher’s date that anchors it, on a page a crawler can reach. The rest of the record got the same treatment, on its own page, every year since 2017.

Get it published, then get it readable: a date no machine can find defends nothing.