Do AI assistants search Google and Bing for me?

Many do, and their searches show up in your search data. Across six months of Bing data for 41 sites, I found 130 long queries with machine fingerprints, such as instructions to return a JSON array or year and month tokens, and one phrase repeated in 24 variants with endings like step by step, examples and best practices. They were seen 492 times and clicked 16 times.

Are long search queries worth targeting?

The human ones are, but not one by one. Queries of ten words or more were under 1% of impressions on my sites but were clicked at 14.78% to 19.87%, against 2.76% for queries of one to three words. Most were seen only once or twice, so build pages that answer the underlying question with labelled sections, rather than a page per query.

How do I optimize a page for AI query fan out?

Answer the core question first, then add clearly labelled sections for the endings assistants add. Machine generated searches repeat suffixes such as step by step, examples, best practices and latest, often with a date, so a page with those sections and an honest last checked date covers most of them. Use repeated query families as an outline, not as keywords.

The Queries No Human Typed

¶ DATATHE ODD ONES

01A search nobody would type

One of the queries in my Bing data is thirty words long. It asks for a list of tools that monitor two assistants, then instructs the search engine to return only a JSON array with specific keys for each object. No person types that into a search box. It is a prompt, written by software, that ended up in a search index and was matched against one of my pages. Some of your search impressions are not people. They are assistants searching on a person's behalf.

A family of 24

Once I noticed that one, I went looking for others across all 41 sites I have verified, about six months of data, 695,483 impressions. I found a whole family of them: 24 versions of the same long phrase about landing pages and paywalls, each with a year and month attached and a different ending. Step by step. Examples. Best practices. Architecture. Latest. Reddit discussion. That is not 24 people. That is one question being expanded by a machine into the searches it thinks will cover it.

// LONGER QUERIES, RARER, CLICKED MORE~~BING · 41 SITES · 695,483 IMPRESSIONS~~

LENGTH

SHARE OF IMPRESSIONS

CLICK RATE

1 to 3 words

71.4%

2.76%

4 to 6

25.7%

4.46%

7 to 9

1.8%

8.24%

10 to 14

0.7%

14.78%

15 or more

0.3%

19.87%

24variants of one phrase, differing only by a date token and a suffix such as step by step or best practices

130long queries with machine fingerprints, seen 492 times, clicked 16 times

Long queries are rare and mostly seen once. Some were written by people. Some, clearly, were not.

¶ DATAWHAT LONG QUERIES ARE

02Two populations in the long tail

The people

Most long queries are still human, and they are the best traffic anyone gets. Queries of ten words or more made up under 1% of impressions, but they were clicked at 14.78% to 19.87%, five to seven times the rate of short queries. A person who types a whole sentence knows exactly what they want, and when a result matches, they click. Those are often the questions where a computed or specific answer is needed, the pattern from calculators and articles.

The machines

The other population looks different. I found 130 long queries carrying obvious machine fingerprints: instructions about output format, pasted web addresses, year and month tokens in the middle of a phrase. They were seen 492 times and clicked 16 times. That fits what an assistant does when it researches a question: it writes several searches, reads the results, and answers without a visit, the same fan out that one question asked of five engines showed from the other side. A search that ends in step by step, examples, best practices and latest, all at once, is a machine making sure it has not missed anything.

I cannot prove that any single query came from a machine, and I would not want to overstate it. The signatures are the evidence. But 87.9% of long queries were seen only once or twice, against 71.5% of short ones, which is what you would expect if many of them were generated rather than typed.

EXTRACTED — THE SENTENCE THIS ENTRY EXISTS FOR

Write for the question, and the searches the machine invents to answer it.

¶ DATAWHAT TO DO WITH IT

03Answer the expansions, not every variant

Build for the suffixes

The useful thing about machine searches is that they are predictable. The endings repeat: step by step, examples, best practices, latest, and a date. A page that answers the core question and then has a clearly labelled step by step section, real examples, a short list of practices, and a line stating when it was last checked covers most of what the fan out will look for. That structure also helps the opening of the page do its job, which matters because the first thirty percent of a page carries most of what gets quoted. You do not need a page for every variant. You need one page with a section for every ending.

Do not chase them

The opposite mistake is writing pages for these strings. They are seen once, never repeat, and are almost never clicked. Treat them as a view of how assistants search your topic, not as keywords. Look for the families: the same phrase with several endings tells you which sub questions an assistant thinks belong to a topic. A query family is a free outline of the page an assistant wishes existed. And date your pages honestly, because a phrase searched with a month attached is asking for something current.

To find them in your own data, export your queries from Bing Webmaster Tools, which keeps about six months, sort them by number of words and read the longest hundred. Look for three fingerprints: an instruction about output, a pasted web address, and a date token in the middle of a phrase. Then group the rest by their first eight words. A group with several different endings is a fan out family, and it shows you how machines are researching your subject. It takes about twenty minutes.

Tomorrow: back to the record, with the same rule as every entry, real numbers and the method shown.