From 142k to 838k Google impressions: one consultant’s AI SEO rebuild
On 8 October, a saved ChatGPT answer to the question “who is the best AI marketing consultant in the UK for a small business” opened with: “My first choice: Lilach Bullock.”
That result sits on top of a broader rebuild. Lilach’s site went from 142,439 Google impressions in February to 838,237 in the 28 days ending 6 October. Clicks went from 937 to 7,087. Bing’s AI Performance report showed roughly 328,700 citations over three months, averaging about 11,200 a day during 1-8 October. Sponsored and partnership enquiry conversations grew from 186 in the first nine days of September to 335 in the first nine days of October, roughly 80% growth.
And in between all of that, she crashed her website, moved hosts, and discovered her missing contact form enquiries had been sitting in Gmail’s Trash the whole time. This is the full breakdown.
The Numbers in Full
The two comparison periods are February (142,439 impressions, 937 clicks) versus the 28 days ending 6 October (838,237 impressions, 7,087 clicks). Lilach notes her goal was one million impressions. At 838,237, she’s at roughly 84% of it.
For a shorter comparison, clicks rose from 4,579 during 7 August to 3 September to 7,087 during 9 September to 6 October, about 55% more in the later period.
GA4 recorded 23,228 sessions in the 28 days from 12 September to 9 October, against a stated goal of 20,000. Within those sessions, 3,730 came from sources attributed to AI assistants. ChatGPT became the third-largest source of visits, ahead of Bing, sending 3,461 sessions during that window.
Bing’s AI Performance report tracked average pages cited per day rising from 60 during 11-31 July to 653 during 1-8 October. Across the full report, 1,319 different pages were cited for 1,468 grounding queries. The daily peak reached about 13,500 citations. Lilach notes the cause of a sharp jump on 16 September is unproven, so she does not attribute it to a specific change.
️ What the SEO Work Actually Involved
The work focused on pages that were already getting impressions, not on creating content about untouched subjects. Lilach used Google Search Console to find pages appearing for useful queries but either answering them badly or failing to connect to a relevant service.
Titles and answer blocks
During 2-5 September, she rewrote titles to match the questions already bringing impressions to those pages. On 8 September, she added short answer blocks to top pages: two or three direct sentences beneath the relevant heading, before the supporting detail. The test she describes is simple: read a heading and the paragraph beneath it in isolation. Does that give a useful answer, or does the reader have to work through four more paragraphs to find it?
Internal links and content connections
She added internal links so related articles connected to the relevant service page, rather than leaving the reader to find it themselves. The framing she uses: if an article answers a question but never connects to the service that solves the underlying problem, the reader has to do the detective work.
Using citation data to prioritise
Bing’s AI Performance report lists the search queries behind citations. On 2 September, Lilach used that data to add sections answering specific questions where her site was already being cited. She notes the pinned-tweet milestones query was at 11.72% citation share before she added a dedicated section. She also added an llms.txt file, checked that AI crawlers could access the site, and created a dedicated AI information page about her business.
How She Crashed the Server
On Sunday 27 September, the site went down from about 9am until just before 11am, with further dips that afternoon.
The cause: nine AI sessions uploading articles and images simultaneously on a server with two CPUs. One session also saved a site-wide template at the same time. Fallback code designed to handle missing pages kept sending requests back to WordPress, adding load to an already struggling server.
The fix was a shared publishing queue. Now only one tool can write to the live site at a time. Site-wide changes including templates and menus are restricted to 11pm-7am. Hourly, daily, and monthly limits are in place, plus an uptime check every five minutes that pauses the tools when the site slows down.
The Hosting Move
On 4 October, Lilach moved from Flywheel to 20i’s X Large plan: eight cores, 16GB of RAM, behind Cloudflare.
She tested a smaller plan first. It returned a load time of 4.35 seconds, with 20 pages taking over six seconds, so she upgraded before switching. The larger plan handled 16 simultaneous requests across 320 uncached pages with zero errors and a median response of 1.73 seconds.
⚠️ What Broke After the Move
Several problems surfaced after migration. The most time-consuming one had nothing to do with the host:
- Missing form emails: An old Gmail filter was sending contact form submissions straight to Trash. Lilach had been looking for a hosting problem. The messages had been arriving the whole time.
- Malware flag: Two already-quarantined files triggered a flag that disabled email and scheduled jobs. She removed them, got a clean rescan, and had support lift the flag.
- Missing index file: The site’s main index file disappeared briefly. Restoring it brought the site back. The cause was not established.
- Blocked scripts: The firewall blocked her own scripts. She updated 329 of them to identify themselves clearly.
- Cloud agents writing to the old server: For a few hours after the move, agents kept targeting the previous server. She changed their destination and reran the affected work.
- Vanished scheduled jobs: Scheduled jobs disappeared when an AI session ended on 6 October. Dated work now lives in a calendar file with a lasting scheduler that picks up missed jobs.
The ChatGPT Recommendation
On 8 October, both ChatGPT and Perplexity recommended Lilach in saved web-interface answers to the same query about hiring an AI marketing consultant for a small business in the UK. ChatGPT’s answer opened with “My first choice: Lilach Bullock” and described her as combining “long-term marketing experience with practical AI implementation.” Perplexity cited her guide to AI consultants for small businesses and her comparison of AI marketing consultants.
The pages ChatGPT linked to were her homepage and her guide to AI consultants for UK small businesses. Both pages state her name, the work she does, and who it’s for. She notes she cannot isolate which individual change earned the recommendation, and that AI answers vary between runs. The 8 October answers are dated results.
The Replication Framework
Lilach describes how she would apply the same sequence to any service business. The bookkeeping example she uses is illustrative, but the logic transfers:
- Pick the service you want more enquiries for. Make the page specific enough that a buyer can recognise themselves in it. “Helping your business thrive” could describe almost anyone. “Monthly bookkeeping for freelancers who want their accounts up to date” gives someone something to buy.
- Find the questions that come before a sale. Use search data and what prospective customers already ask you.
- Improve the answers and the route to the service. Each change should help someone understand the offer or reach it.
- Save the starting figures and dated AI answers. Check the same pages and questions again after the work.
- Follow the enquiry through. Confirm the form works and the message lands somewhere you will actually see it.
If tools are making the changes, add a shared publishing queue and a way to pause the work when load spikes. Lilach describes testing the enquiry form end to end as non-negotiable, having spent time blaming a host for a Gmail filter she had set herself.
Output Volume After the Fixes
The scoreboard recorded 4,795 successful write requests across 4-10 October, about 685 a day, versus a September baseline of 288 a day. By 9pm on 10 October, that day’s count alone reached 1,124. The content engine runs on Cloudflare and is set to about 50 posts a day. The overnight run from 10 to 11 October made another 265 changes.
Lilach notes the change counter includes edits and uploads, so several operations can belong to one page. Page-creation requests and published pages are tracked separately.
When This Works and When It Does Not
The results here span months of consistent work on titles, internal links, content improvements, and technical fixes, not a single campaign. Lilach is explicit that she cannot attribute the full growth to any one hosting move or one busy week. The 16 September citation spike in Bing’s data has no proven cause.
The approach is most directly replicable for service businesses with existing content that is already appearing in search but not converting, where clearer answers, better page structure, and tighter internal linking can move a reader from finding you to contacting you. It is less directly applicable to businesses with no existing content or search presence, where the baseline work takes longer before data is available to act on.
I’m founder of BizStack at Cagri Sarigoz LLC and a passionate advocate for entrepreneurs.
With over 14 years in tech, marketing, and AI, including my role as Head of SEO at CitizenShipper and co-founder of TaleBot at Intale AI, I’m dedicated to sharing genuine, useful product insights and tips.
At BizStack, I aim to cut through the digital noise to provide clear, actionable advice.
And more than all else, I’m a father to a (always) little girl and a husband.
Contact me at [email protected] for assistance.
