Field note

AI Overviews are eating informational queries - and leaving the money ones alone

By James Frost, Founder, WARDORX

Last reviewed

10 min read

The queries that convert are still a blue-link fight. Here's how that should change where your content budget goes.

The note

The short version

The dividing line is not informational versus commercial. It is whether one answer serves everybody who asks. A model can answer why a furnace short cycles, because the answer is the same in every city; it cannot answer who should fix yours, because it does not hold the input that decides. Queries whose answer is a fact get summarised. Queries whose answer is a decision about a specific person still send the click - which is why service and location pages, the map pack and on-page conversion are worth about seven of every ten hours, and why your cost pages are the ones to watch, not your service pages.

The line is not informational versus commercial

The split is easy to see for yourself. Search 'why is my furnace short cycling' and you will usually get an AI Overview. Search 'furnace repair near me' and you get a map pack and blue links. Everyone has noticed this, and almost everyone has drawn the wrong line through it - informational gets summarised, commercial does not - which is a description of the symptom rather than the cause, and it fails on the queries that matter most.

The actual line is whether one answer serves everybody who asks. Short cycling has a cause list that is identical in Calgary, Leduc and Toronto, so a model that has read enough of the corpus can produce it, and producing it is genuinely better for the person asking than ten links that each produce it. Who should repair your furnace has a different correct answer for every person who asks, decided by where they are standing, who is licensed there, who answers the phone tonight and who has a track record with their neighbours. The model does not hold those inputs. Google does, which is why it hands the query to the map pack instead.

That distinction matters because it predicts rather than describes. Intent taxonomy tells you what already happened to a query. The one-answer test tells you what is going to happen to a query you have not published for yet.

The test: would two neighbours be well served by the same answer?

Take any query you rank for and ask whether two people in different neighbourhoods, asking the same words on the same day, would both be well served by one identical answer. If yes, the query is a candidate for summarisation, and you should assume it will be summarised eventually whether it is today or not. If no, it is structurally safe for as long as the answer depends on inputs the model does not hold.

Run that down a keyword list and it sorts in an afternoon. What comes out is not the split most contractors expect. Diagnostic questions, maintenance intervals, code and permit explanations, product comparisons and how-long-does-it-last questions all land on the summarised side. Emergency and same-day queries, hire-and-book queries, and anything carrying a neighbourhood name land on the safe side. That much is unsurprising.

The boundary moves, and it moves one way

The surprise is in the middle, and the middle is where money already is. 'How much does furnace repair cost in Calgary' is a fact-shaped question wearing a place name. It looks local, so it gets filed with the safe queries, but the answer is a range, ranges are exactly what a model assembles well, and the local modifier narrows the corpus rather than changing the kind of answer required. That is a query in the process of crossing.

So the front line is not your service pages. It is your cost pages - the ones a lot of contractors built in the last two years precisely because they pulled traffic. They will keep pulling impressions and they will stop pulling the same clicks, and the useful response is not to abandon them. It is to put something on them that a summary cannot carry: your actual numbers, dated, with what changes them and what a specific job in a specific city came to. A generated range is an average of everything published. A page that says what this work costs here, this year, and shows the arithmetic, is the page the summary has to cite in order to be worth anything.

Watch the same movement coming for comparison queries and for anything phrased as which is better. The direction of travel has been consistent: as soon as an answer can be assembled correctly without knowing who is asking, it gets assembled.

What informational content is now for

If a post can no longer be counted on for clicks, the honest question is why publish it. The answer is that its job changed rather than disappeared, and the new job is worth more per dollar than the old one for a small local business - it is just measured somewhere else.

An informational page now buys three things. It buys presence in the corpus a model reaches into, so your name appears beside the topic when the topic comes up. It buys entity strength - the accumulation of consistent, specific, checkable statements attached to one business, which is what lets a retrieval system treat you as a thing it knows rather than a string it matched. And it buys the citation itself, which puts your name in front of a person at the moment they are still deciding whether this is a job for them or for a professional.

None of those three is a session. All of them are real. The failure is not that informational content stopped working; it is that it stopped working in the column people read.

Why the reports say to cut the pages doing the work

The chain from an informational post to money used to be one step: post ranks, person clicks, person calls. It is now three. The post gets cited. Some fraction of the people who see the citation search your name later, days or weeks later, because a name seen at the research stage is what they reach for when they are ready. That branded search converts far better than anything non-branded, because someone typing your business name has already made the decision that everything else in your funnel is trying to make for them.

Three steps, a lag of weeks, and only the last one shows up as a conversion in the report. So the post looks like a failure in exactly the view a contractor reads - sessions down, conversions from that page at zero - while it is quietly feeding the channel with the best conversion rate on the site. Cut it and the branded search that was building on it stops building, six weeks after you stop being able to connect the two events.

The way out is to change what you measure on those pages before you change the pages. Track branded search volume as a line of its own, monthly, and read it against what you published two months earlier rather than what you published this month. Track AI impressions on the pages you meant to be cited. Neither is a conversion metric and neither should be dressed up as one - but between them they tell you whether the informational half of the budget is doing anything, which a session count no longer does.

How a citable page is built differently to a rankable one

The craft changed too, and this is the part that is easy to get wrong because the old craft was good craft. A page built to rank builds an argument: it opens with context, develops across sections and lands its conclusion at the end, which is how you keep a reader who is scrolling. A page built to be cited has to survive being cut into pieces, because that is what happens to it. A retrieval model lifts a passage. It does not read to the end.

So the unit of a citable page is the self-contained paragraph: one claim, its qualification, and the number or condition that makes it checkable, all inside a block that still makes sense with nothing above or below it. Headings become questions rather than labels, because a heading is the handle a passage gets selected by. The answer goes in the first sentence under the heading rather than the last. Anything load-bearing gets a date and a source attached to it in the same paragraph, not in a footnote, because the footnote does not travel with the quote.

This costs nothing in readability for humans, which is the pleasant surprise. Answering the question in the first sentence and then explaining it is how good technical writing worked anyway. What it costs is the pleasure of a slow build, which was always more fun to write than to read.

Where the seventy and the thirty actually go

We weight retainer hours roughly seventy to thirty against informational content, and it is worth saying what each half buys rather than leaving it as a ratio.

The seventy goes to service-by-location pages, the map pack and conversion rate on the page itself - the three places where the answer still depends on who is asking, and therefore the three places a summary cannot take from you. Within that seventy, the largest single line is usually not new pages at all. It is fixing the last join: the form, the phone, the thing that happens after someone has decided to hire you, which is the step with the worst numbers on most contractor sites and the one nobody is reporting on.

The thirty goes to informational work built to be cited rather than to rank, aimed at the questions that come immediately before a job - the diagnostic ones a homeowner asks at eleven at night with a cold house. Not because those posts will produce leads, but because that is the last moment before the decision, and being the name attached to the explanation is worth more than being the tenth link under it.

The one edit we made to every informational page

One change, applied everywhere, did more than any of the rest: every section heading became a question, and the sentence directly under it became the complete answer to that question - stated outright, no set-up, no dependency on the paragraph above.

It is a mechanical edit and it takes about twenty minutes per page. What it does is make every section independently quotable, which turns one page with one chance of being cited into a page with eight. It also exposes, brutally, any section that did not actually have an answer in it. Several of ours did not, and those sections were deleted rather than rewritten, which is the other thing the edit is good for.

Check it on your own site

  1. Sort your keyword list with the two-neighbours test

    Export the queries you get impressions for. For each one, ask whether two people in different neighbourhoods would both be well served by one identical answer. Yes goes in the summarised column, no goes in the safe column, and anything you argue about for more than ten seconds goes in a third column called crossing. The third column is the one to act on.

  2. Reinforce the cost pages before they cross

    Take everything in the crossing column - usually cost, comparison and how-long questions with a place name attached - and add what a generated answer cannot have: your own dated figures, the conditions that move them, and one real job in one real city with the arithmetic shown. A summary that has to cite you to be useful is a better outcome than a ranking you were going to lose.

  3. Separate the two budgets and label them honestly

    Split the content line in your plan into lead-generating work and citation work, and stop reporting them in the same column. Lead work is judged on form submissions and calls. Citation work is judged on AI impressions for that page and on branded search volume two months later. Mixing them is how the second kind gets cut in a quarterly review.

  4. Put branded search on the monthly report as its own line

    In Search Console, filter queries containing your business name and record the monthly total as a standing line item. Read it against what you published eight weeks earlier, not against this month's work. It is the only readily available proxy for whether being cited is reaching anybody, and it is free.

  5. Run the heading-to-question edit across the informational archive

    Rewrite every section heading as the question that section answers, then make the first sentence beneath it the whole answer with no run-up. Twenty minutes a page. Any section you cannot write a first-sentence answer for is a section without an answer in it - delete it rather than padding it, and the page gets stronger for being shorter.

Questions this raises
Do AI Overviews appear on 'near me' searches for contractors?
Rarely, and when they do they sit above a local pack rather than replacing it. The reason is structural: the correct answer to a near-me query depends on the searcher's location, licensing in that area and who is currently available, and none of those are inputs a generated summary holds. Google routes the query to the map pack because the map pack is the system that does hold them.
Which of my pages is most at risk of being summarised next?
Your cost pages. A question like 'how much does furnace repair cost in Calgary' looks local because of the place name, but the answer it wants is a range, and a range is exactly what a model assembles well from published material. The place name narrows the corpus rather than changing the kind of answer needed, which is what makes it a crossing query rather than a safe one.
If informational posts do not bring leads any more, should I stop writing them?
No, but you should stop measuring them by leads. They now buy three things - presence in the corpus a model retrieves from, entity strength for your business as a known thing rather than a matched string, and a citation in front of someone at the research stage. The return arrives later as branded search, which converts better than any non-branded traffic you buy.
How long is the lag between being cited and seeing anything from it?
Weeks rather than days. The mechanism is that someone sees your name while researching, does not act, and searches for you when they are ready - which for home services is usually when something breaks. Read branded search volume against what you published roughly two months earlier. Reading it against the current month will make the work look like it does nothing.
Does being cited in an AI Overview send traffic?
Some, but far less than the equivalent ranking used to, and treating citation as a traffic channel is the mistake that makes the whole activity look worthless. Treat it as brand distribution with a measurable second-order effect. The number that moves is branded search, not sessions on the cited page.
What is the seventy-thirty split and where did the numbers come from?
Roughly seven of every ten retainer hours go to service-by-location pages, the local pack and on-page conversion, and three go to informational content built to be cited. The ratio came from where leads actually originate on the contractor sites we manage, not from a study. It moves for a business with an unusually strong local position already, where the informational half is worth more.
How do I write a page so it can be cited rather than just ranked?
Make every section survive being cut out of the page. One claim per paragraph, with its qualification and its checkable number in the same block. Headings as questions, because a heading is the handle a passage gets selected by. The answer in the first sentence under the heading rather than the last. Dates and sources inline, because a footnote does not travel with the quote.
Is there any way to see which of my pages AI surfaces are actually using?
Partly. Search Console's generative AI performance report gives page-level impressions for AI Overviews and AI Mode, so you can see which pages were reached for. It gives no clicks, no position and no queries, so it tells you which page was used and not what question it answered. Export it monthly - the data is interface-only and has no history behind the rollout.
What to do about it
The money queries in one trade

Which HVAC searches still resolve to a map pack and blue links, and what the pages behind them need to carry.

Where the 70/30 split costs out

What informational and transactional work each buy at a given monthly spend, with the market ranges beside ours.

Bring the site you have. Leave with a clear next step.

Book 30 minutes with James. We will look at the current site, the market you want to win, and whether the right move is a build, SEO, or no project at all.

Book a call - 30 min

live availability
Loading available times

Times shown in your local zone. You'll get the audit doc before the call, not after.