Half the searches reaching us are full sentences. A lot of them aren't people

You will have been told that search is becoming conversational — that people have stopped typing three words and started asking questions, and that your website needs to change accordingly.

Rather than repeat that, I counted. This is every search query Google showed us for our own site over ninety days, sorted and classified by hand. It is a small sample from a small site and I will keep saying so.

The short version

Exactly half the individual queries we could see were written as sentences or questions rather than keywords. That part of the story is true.

But they were half the variety and only about a quarter of the volume. And when I looked at the longest, most conversational-looking ones, a good number of them were plainly not people at all.

What we actually looked at

Ninety days to 30 August 2026, our own site, pulled from the Search Console API. Over that period the property recorded 3,272 impressions and 43 clicks, an average position of 13.6.

Google itemised 140 individual queries for us, covering 990 of those impressions.

I classified a query as conversational if it ran to six words or more, or began with a question word, or ended in a question mark. That threshold is mine and it is arbitrary; a different line would give a different number. Here is what it produced:

• 70 conversational — 267 impressions, average position 19.6
• 70 keyword-style — 723 impressions, average position 23.0

An exact 50/50 split of queries, and roughly 27/73 by impressions. So the long questions really are there, in numbers, but each one is rare. The short keyword queries are fewer in kind and far more often typed.

The real ones are lovely to read. “Which ai agents work inside slack or teams like a real coworker?” — twelve words, with the question mark. “How can a marketing team update the website without waiting on developers”. And the one that produced yesterday's post, at fifteen words: “is there a way to create a website that never goes down or needs updates?”

You can hear an actual person in each of those. Which is what made the rest stand out.

Then I looked at who was asking

Three things in that list are not somebody at a keyboard.

1. The same query, reworded twenty-four times. We had 24 near-identical variants of one query about EU AI Act deepfake disclosure obligations — strings like “eu ai act transparency obligations ai generated content deepfakes august 2026 official”, then the same words in a different order, then again with “obligation” singular, and so on, across 39 impressions. People do not do that. Automated monitoring does.

2. Search operators. One query was “marketing automation” in quotes followed by a run of -site: exclusions knocking out Reddit, X, YouTube, Yelp, TripAdvisor and others. That is somebody's scraping or monitoring tool, configured once and left running. Another was a boolean lookup of our registered office address, brackets and all — a data broker or a verification service checking we exist.

3. They never click. Which brings us to the most interesting number in the whole exercise.

All 140 queries. Zero clicks.

Every itemised query we could see — conversational and keyword alike, 990 impressions between them — produced not one click in ninety days.

Meanwhile the property recorded 43. So roughly seven in ten impressions, and every single click, came from queries that were not in the list we could read.

I want to be careful here. I am describing that gap, not explaining it. I went looking for Google's own documentation of how the unlisted remainder is handled and could not find it on the pages I checked, so I am not going to guess at the mechanism — you will find plenty of confident explanations elsewhere and I cannot verify any of them.

What I can say is the practical bit: the query list is not a picture of your traffic. It is a sample, it skews to the rare and unclicked, and if you plan your content around it you are planning around the part of your audience that did not visit.

What argues the other way

One small site is not evidence about search. We are a young site with modest traffic in a technical niche, which is exactly the sort of site that attracts monitoring tools and exactly the sort with too few searches per query to be typical. A busy plumber in Leeds would see a completely different shape. Take the method, not the numbers.

My bot-spotting is inference, not proof. I cannot see who is behind a query. A determined researcher could plausibly type eleven variants of the same AI Act question, and someone genuinely does use search operators. I think the pattern is clear, but I am reading tea leaves and you should know that.

And conversational search may still be growing. Ninety days is one window, not a trend. If people are shifting towards asking questions, a single snapshot cannot show it, and I have not claimed otherwise.

What to do about it on Monday

• Go and read your own query list before you buy anything sold as “AI search optimisation”. Twenty minutes, free, and more informative than any article about the trend — including this one.
• Look for the sentences. If somebody asked your site a whole question, that question is a page you have not written, or a page whose title does not admit it answers it.
• Discount the permutations and the operator queries. They inflate your impression count and will never buy anything.
• Do not chase the trend as a project. Answering the real question plainly, near the top of the page, in the customer's words, works for a person skimming, a search engine matching and a model summarising. That is the whole optimisation.

Sources, method and limits. All figures are from our own Google Search Console property (sc-domain:lolasquared.com), pulled via the Search Analytics API on 31 August 2026 for the 90 days from 2 June to 30 August 2026: 3,272 impressions, 43 clicks, CTR 1.31%, average position 13.6 at property level; 140 itemised query rows covering 990 impressions and 0 clicks. Search Console data lags by a day or two, which is why the window ends on 30 August. The conversational / keyword classification is our own, defined as six or more words, or beginning with a question word, or containing a question mark, and applied by script to all 140 rows; positions quoted for each group are weighted by impressions. The near-duplicate families were grouped by token overlap rather than by hand. We did not verify why the itemised rows do not sum to the property total — we checked Google's Search Console help pages on impressions and on the Performance report and found no statement of the mechanism, so the post describes the arithmetic and stops there. This is one small site over one quarter and should not be read as a finding about search in general.