A local services business in 11 U.S. cities relaunched its website on September 9, 2026, with every page rewritten using Herringbone GEO Tools. Here is what happened to its Google rankings (SEO) and to how often Google’s AI answers recommend and cite it (GEO).
Can the patterns we found help a small business get recommended by AI?
We build tools that find patterns in what search engines and AI answers reward. We wanted to know whether those patterns, applied to a real small business, could lift its Google rankings and, more importantly, get it recommended by AI. So we tested it on a business that serves customers in 11 U.S. cities.
Why this business. We picked a business with a lot of room to grow. People in its cities were searching for what it offers, but the site barely reached them. In the two weeks before its first changes, searches that didn’t name the business brought about two clicks a day. AI answers rarely recommended the business by name. It served customers in many cities, so we could see whether the same approach held up across different markets.
Where it started: Google showed the site, but hardly anyone clicked.
Before September 9, the site had basic city pages, live since July 1. They had started to earn some searches, but they were thin, and search was still a small part of how people found the business.
- Few high-volume rankings.In August, the site ranked in the top 3 for none of its tracked high-volume searches, and in the top 10 for just six. Its top-3 spots on smaller searches brought only 126 clicks.
- Search was a minor source.Of about 126,000 recorded visits since January 2025, 13% came from search. Direct, social and paid visits made up most of the rest.
- Pages too thin to rank.Location pages ran 130 to 250 words, with incomplete meta tags and headings, image descriptions that named the wrong city, and copy errors.
So the question was not just how to rank higher, but what an AI looks for when it decides which business to recommend.
Three things AI looks for before it recommends a business.
To find out, the tools study the searches an AI runs behind the scenes to put together an answer and a recommendation, along with the questions customers most often ask. Together they show which topics a page has to cover and how it should be structured. Three things stood out.
Semantic directions: cover every angle of the search in one place
A search like “[service] lessons in [city]” is not one question. It splits into several directions, the angles an AI answer tends to cover. The tools run three checks on a page.
- They split the search into directions. This one becomes price, what is included, who it is for, where it happens and who teaches. Most searches have three to six.
- They grade every sentence. For each direction, a sentence is either not addressed, a general claim (“affordable lessons”), or specific evidence (“lessons start at [price] for a 2-hour session”).
- They find the best passage. That is the one stretch of the page, a few hundred words, that covers the most directions with specific evidence.
The passage is what counts, because an AI quotes passages, not whole pages. It lifts a few hundred words from the one that covers the most and drops the rest. A page can mention every direction somewhere and still lose if no single passage covers them together. SEO gets the page read. The passage decides whether the business is used.
Fan-outs: answer the questions AI asks behind the question
When someone asks an AI a question, it often breaks it into several related sub-questions before it answers. A page that already answers those sub-questions gives the AI more to use. This is why the site gained 55 answered questions.
Credible and specific: say real, local things
AI prefers content it can trust and attribute to a specific place or source. The thin pages became full location pages with local detail such as neighborhoods and venues, and the generic image descriptions became specific ones.
Those three ideas turned into two steps. Step one is the SEO foundations: fixing what kept Google from reading and ranking each page. Step two is the GEO layer: adding the specific content an AI needs to recommend the business.
Step one: make every page readable to Google.
Most of what the audits flagged were SEO foundations. They matter because a page with a vague title, a missing heading or a wrong canonical tag is hard for any search engine, or any AI that draws on search, to read. The tools audited every page against Google’s own search data. Each recommendation names the exact current text, the exact replacement and the search it targets. Every change was applied. Examples use placeholders for the city and service.
Page titles: match the words people search
The audit found a city page ranking 4th for “learn [service] in [city]” but earning no clicks. Another sat on page 2 for the plain “[service] [city]” search. Both titles were generic and identical in structure across cities.
The new title covers both “lessons” and “classes”, the two words searchers use, and puts the city early. Some titles also gained the nearby area or regional name people use.
Meta descriptions: say what you get and where
Image descriptions: say what the picture shows
The tools reviewed the alt text on every image across 15 pages. About 35 were already good, about 69 could be improved, and about 4 decorative images should be skipped by screen readers.
Technical fixes and copy errors
- Canonical tag.The homepage’s canonical pointed at a slightly different address, which tells Google the page is a duplicate. Fixed so it points to itself.
- Missing headings.A page with no top-level heading could not tell Google its topic. One was added.
- Pages that should not rank.The cart page was set not to appear in search results.
- Structured data.Site-wide markup warnings were fixed, including absolute image addresses.
- Copy errors.The tools also caught typos, a wrong city name, a broken link and all-caps testimonials: 9 in all.
These fixes get a page found and ranked. They do not, on their own, give an AI a reason to recommend the business. That takes the second layer.
Step two: answer the questions AI asks, in specific local terms.
Step one helps both Google and AI. Step two adds what AI needs on top. To be recommended in an AI answer, the page also has to hold the specific things the AI is looking for. For this business, three changes carried that: answering the follow-up questions an AI asks behind the original question, stating the specific details a good answer has to cover, and naming the local places the business serves.
1. Answer the questions behind the question
When someone asks an AI about “[service] lessons in [city]”, it fans out into related questions before it answers. Every location page now answers the ones that come up, in plain question-and-answer form, marked up so Google can read it. That is 5 questions on each of 11 pages, 55 in all.
2. State the specifics a good answer needs
The direction analysis showed what a strong answer to “[service] lessons in [city]” has to cover. Each location page now states those details outright, in one section, instead of leaving them to a contact form.
- Price.A starting price for a private session and ticket prices for public events, with what each includes and what changes it.
- Length, format and outcome.How long a session runs, what happens in it step by step, and what a beginner can do by the end.
- Options.Beginner lessons, guided play and open play at public events, and custom corporate events.
- What is included.All equipment provided, so nothing to bring. For corporate events, what the business supplies and what the client provides.
- Who it is for.Beginners, friend groups, families, celebrations and teams, with no experience needed.
- Who teaches, and why trust them.Instructor profiles with real backgrounds and years of teaching experience, plus student reviews.
3. Say where, in specific local terms
AI favors content it can attach to a real place. The thin pages talked about “the surrounding area”. The rebuilt ones name the neighborhoods, the nearby suburbs and the venues, and have a “where we teach” section for each city. Image descriptions name the city too.
Other local touches: a closing heading that adds the nearby suburbs, an origin-story paragraph on the home market’s page, and image descriptions such as “Instructor teaching a beginner lesson at a public event in [City]”.
Supporting structure
Each location page also grew from 130 to 250 words to 1,150 to 1,650, organized into labelled sections so each topic is easy to find. The site also gained new pages and a blog covering topics people search for that it hadn’t addressed before.
Every change we made, at a glance.
Step one: the foundations
- Title tags.All 16 pages.
- Meta descriptions.All 16 pages.
- H1 headings.One keyword heading on each of the 11 location pages.
- Image descriptions.125 images, now naming the subject and the city.
- Copy and link errors.9 found on the old site, 0 on the new.
Step two: the AI layer
- Page content.11 location pages, from 130 to 250 words to 1,150 to 1,650, with a starting price, session details, reviews, and named neighborhoods and venues.
- FAQs.55 questions, 5 on each location page, in a form Google can read.
- New pages and a blog.Covering topics people search for that the site hadn’t addressed before.
Both steps started from the same research: 24 documents with a keyword target for each city, a live audit of every old page, and page outlines with finished copy.
We measured both steps, and four numbers moved.
Step one was aimed at Google rankings (SEO), and step two at AI recommendations (GEO). These are the four headline numbers, two for each. The AI numbers are what the project was built around, so we look at them first.
more clicks a month from Google’s top 3 results for searches that don’t name the business.
126 in August, 642 in October. Google Search Console.popular search terms now in Google’s top 3. In August, the month before launch: none.
Independent rank history, October.of the AI answers Google shows for its core searches recommend the business.
Before the rebuild: rarely, if ever (client observation). Now 544 of 578 answers, from 22 searches in 11 markets, checked on 34 days since Sep 29.of those answers also cite the site in their source links.
548 of 578, checked since Sep 29.One thing to know. The relaunch also moved the site to a new platform, which was the owner’s choice, not one of the tools’ recommendations. Some of the gain may come from that move.
Google’s AI recommends the business 94% of the time, and links to it 95%.
This is what the GEO layer was aimed at. When Google’s AI Overview answers a search for the business’s core service in a city, it usually lists a few businesses. Recommended means the answer names the business. Cited means it links to the site. Before the rebuild, the business was rarely, if ever, recommended. We began checking on September 29, three weeks after the rebuild, and have checked on 34 days since, across 22 searches in 11 markets.
Core searches that show an AI Overview, by day
An AI Overview appeared for 80% of these searches. The business was named in 544 of the 578 that did, and the site was cited in 548. In the last seven days checked (Oct 28 to Nov 9), 120 of 122 AI answers recommended the business, and 120 cited the site. Each point is one day, adding up all 11 markets. Days with no check, including Nov 3 to 7, are skipped.
- AI answers shown (the total each day)
- Of those, name the business (recommended)
- Of those, link to the site (cited)
AI recommendation is one half of visibility. The other is ordinary Google ranking, where the SEO foundations show up. The next three sections follow it from daily clicks to the rankings behind them.
Clicks were already rising. After the rebuild they averaged three times higher.
Now the Google side, starting with the simplest measure: each day’s clicks from people who searched without naming the business, from the start of tracking. On July 1 the business added basic city landing pages, which began earning clicks. The September 9 rebuild turned them into full pages, and the line steepens after.
Clicks per day from searches that don’t name the business
- Each day
- 7-day average
Popular searches: no top-3 results in August, 35 by October.
Clicks follow from where the site ranks. This chart counts how many popular search terms, the phrases people type into Google, the site ranked for in Google’s top 3 and top 10 each month. The data comes from an independent rank-tracking service, and it is the only month-by-month record from before launch.
High-volume search terms in the top 10 and top 3, by month
Estimate from DataForSEO for U.S. English searches, January to October 2026.
- In the top 10
- In the top 3
Searches that name lessons or classes plus a city averaged position 2.1 on Google in the last week of October. Ranking is one half of the picture. The other is how many different searches the site shows up for at all.
The site now shows up for twice as many new searches each week.
Ranking for more terms is what widens the funnel. These charts show how many different searches the site appears for, and how fast that number and the clicks grew.
Organic search visits by month
Jan to Aug 2026 are read from the old site’s traffic chart and are estimates. Sep and Oct are Google Search Console clicks.
- Old site (estimated)
- New site
October visits are 40% above the best old-site month. Old-site months are estimates.
Different searches the site is shown for, per day
Searches that do not name the business, counted each day the site was shown. Search Console, from Jun 24.
- Each day
- 7-day average
Before the rebuild is Jul 7 to Sep 8, when the location pages were live. After is Sep 9 to Nov 7. Searches that don’t name the business. The first card is the growth in different searches shown per day, per week (a best-fit line through each period). The second is the average clicks per day in each period.
Different searches the site is shown for each day
2.2× faster after the rebuild
Average clicks per day
3.3× more clicks a day after the rebuild
Terms in the top 10 and top 3, by month
Google’s own data for searches that do not name the business. Terms are counted by their average position over the month. Searches Google withholds for privacy are not included. The first bar is only the 13 days before July, when the site had very few location pages, so it is a partial month.
- In the top 10
- In the top 3
The rebuild went live on September 9, part way through the September bar.
Clicks from top-3 terms, by month
- Clicks from terms in the top 3
126 clicks in August, 642 in October: the 5× in the headline numbers.
How we got these numbers.
- Search Console.Clicks, searches and pages for the client’s domain, Jun 24 to Nov 7, 2026. The page for the market added Oct 6 is excluded. Brand searches are any that contain the business name. Searches Google withholds for privacy are not included, so true non-brand counts are higher.
- AI Overview tracker.A daily check of Google results for core, near-me and corporate searches in each market. It records whether an AI Overview appears, whether the site is cited in its source list, and whether the business is recommended, meaning its name appears in the answer text. 22 searches (its two core terms in 11 markets), checked daily from Sep 29.
- DataForSEO rank history.Monthly count of terms ranking in the top 3, top 10 and top 100 for the domain, Jan to Oct 2026, U.S. English.
- Old-site analytics.Visits by source, Jan 2025 to Sep 2026, read from the Traffic Sources chart. The monthly values sum to within 3% of the platform’s 16,726 search visits.
- Recommendation documents.The recommendation documents show the structure of the changes. Counts for pages, FAQs, words and images come from the location pages as they stood on Sep 29, 2026, and tags from the live pages, before any later edits.
See what Google’s AI is recommending in your markets
We run the same two steps against your site: the audit that finds what keeps Google from reading each page, and the analysis that shows which topics an AI answer has to cover before it will recommend you. You get the current wording, the replacement, and the search each change targets.
Herringbone GEO Tools · Market Report
11 markets
What Google’s AI answers say when people search your service