We Measured Our Own AI Visibility Every Week for 90 Days (Every Number, Including the Ones That Did Not Move)

The short version

Between late June and mid-September 2026 we ran a weekly AI visibility audit on our own domain, pushleads.com, across four separate AI surfaces: Google AI Overviews, Google AI Mode, ChatGPT and Perplexity. We did not run this on a client site first. We ran it on ourselves, because if the measurement was going to be wrong, we wanted to be the ones who ate it.

Here is what 90 days of weekly data says.

Two of those rows are the whole story. Impressions climbed 62% while clicks stayed flat. And our citation footprint inside AI answers grew 41% over the same window.

That is not a coincidence. That is the trade happening to every local service business in America right now, and almost nobody is measuring it.

Why we ran this on ourselves first

We manage local SEO for around 50 businesses — restoration contractors, HVAC companies, plumbers, roofers, pest control operators, short-term rental managers. Over the last year every single one of them has asked some version of the same question:

“My traffic looks weird. Am I showing up in ChatGPT?”

Until 2026 the honest answer from most agencies was a shrug dressed up as a strategy. Someone would open ChatGPT, type the client’s service and city, screenshot the result and call it a report. That is not measurement. That is a single die roll presented as a probability distribution — AI answers change dramatically between sessions, locations and days.

The problem underneath it is real, though. Three things are happening at once.

1. Google is answering more queries itself

AI Overviews now appear on a large and growing share of commercial and informational searches. When they do, the click that used to come to your site frequently does not.

2. Buying research increasingly starts inside an assistant

People open ChatGPT, Gemini, Claude or Perplexity and ask “who’s the best water damage company near me” or “how much does mold remediation cost” before they ever touch Google. Whatever domains get named and cited in that answer shape the shortlist. Across the home-services category, the share of buyers starting research inside an AI assistant has moved from roughly 6% to around 45% in a single year. The relevant question is not whether that trend is real. It is what the number looks like next year.

3. None of this shows up cleanly in Analytics

AI referral traffic is small, attribution is messy, and the vast majority of AI influence never produces a referral click at all. It produces a memory. So the dashboard looks flat while the actual influence compounds.

It does convert, though. A restoration company in Northwest Arkansas found us through ChatGPT in October of last year. They became a client, and in the first months of work their revenue moved from roughly $450,000 to $650,000. That lead did not look like anything special in Analytics. It came from being in the answer.

So we built the instrument. Then we pointed it at ourselves for three months before putting a single number in front of a client.

What we actually measure: the four AI surfaces

There are four distinct surfaces, and conflating them is the most common mistake in this space. They behave differently, they cite differently, and they need different work to move.

Google AI Overviews (AIO)

The generated answer box at the top of Google results. We track a keyword set — currently 190+ unique keywords for our own domain — and for each one we record whether an AI Overview appeared at all, how many sources it cited, which domains were cited, and whether we were one of them. This is the highest-volume, highest-stakes surface because it sits directly on top of the search demand you already compete for.

One finding here is worth stating early: YouTube is the single biggest driver of AI Overview citations we can identify. Across our full client fleet, YouTube shows up in roughly half of the AI Overviews we track. If you take one tactical thing from this report, take that one.

Google AI Mode

Google’s conversational search surface. We started collecting this in early September 2026. The sample is still thin — 22 prompts in the first full sweep — which is exactly why we are flagging it rather than trumpeting it.

Conversational assistants (ChatGPT, Perplexity)

We run a fixed prompt set of buyer-intent questions a real prospect would type, weekly, and record whether our brand gets named in the answer, whether our domain gets cited as a source, how many competitors get named alongside us, and where we fall in the list. Named and cited are different outcomes with different causes.

AI referral traffic

The sessions that actually land on the site from an AI surface, split by engine, tracked in GA4 and pulled nightly. This is the smallest number in the report and, in most cases, the least important one.

Four surfaces. Weekly cadence. Same prompts and keywords each time so the series stays comparable. That last part sounds obvious and is exactly where most “AI visibility tools” fall apart — they change the sample and then report the noise as a trend.

Finding 1: Impressions up 62%. Clicks flat. This is the AI tax.

This is the number every business owner needs to internalize, and it has nothing to do with whether you like AI.

June’s 287 clicks was a single viral outlier — one piece that got picked up and shared. We are calling it out rather than burying it, because leaving it in without explanation would make the trend look better than it is.

Strip out June and the pattern is stark: impressions grew 62% from April to August. Clicks went from 55 to 54.

We got dramatically more visible. We got zero more clicks.

If you only look at your Analytics dashboard, that period looks like failure or, at best, stagnation. Traffic flat, effort up. Plenty of business owners have fired their agency over a chart that looked exactly like that one.

But the impression growth is real. Those are real searches where Google decided our page was relevant enough to surface. The click did not come because, increasingly, there is no click to give. Google answered the question in the Overview. The user got what they needed and moved on.

This is the central measurement problem of 2026. The old scoreboard — sessions, clicks, bounce rate — now systematically undercounts the work. You need a second scoreboard that tracks whether you are inside the answer, not just whether someone came to read your version of it.

Finding 2: We are the #2 most-cited domain in our own AI Overview set

This is the headline result, and it surprised us.

Across the last three weekly sweeps (August 30, September 6, September 13) we recorded every domain cited by every AI Overview that appeared for our tracked keywords, then counted.

Further down: Coursera (23), BlueCorona (19), Semrush’s agency directory (18), Wikipedia (18), WebFX (18), OuterBox (17), Semrush (15), Expertise.com (13).

We are cited more often than every competing agency in our vertical. We are cited more often than Wikipedia. The only thing ahead of us is YouTube, and YouTube is not a domain in the normal sense — it is the aggregate of every video on the internet.

Two honest caveats.

First, this is our keyword set. These are keywords we chose to track because they matter to our business. Any domain measured against a keyword set chosen for its own relevance will look good. This number is not “PushLeads is the #2 SEO agency in the world.” It is “within the specific set of queries that describe our business, we have built the deepest citation footprint.”

Second, that is still the whole point. Nobody needs to win every query on the internet. You need to win the 190 queries that describe what you sell in the place you sell it. That is the job.

The citation trend

Absolute citations grew 41% in three weeks (17 → 24). The rate barely moved, because we expanded the tracked keyword set at the same time — we added 39 keywords, most of them harder and newer.

That distinction matters and most reports would hide it. Growing your absolute footprint while holding your rate steady against a harder keyword set is a genuinely good result. Claiming the rate “grew 6%” would be dishonest framing of a 1.2-point move inside sampling noise.

What we actually get cited for

The September 13 sweep put pushleads.com inside AI Overviews for queries including: disaster restoration seo, restoration seo, restoration company seo lead generation, how to start a water restoration company, local seo firm, local seo service, professional seo service, search engine marketing expert, seo companies, seo firms, seo service, small business seo services, pest control gbp optimization, pest control keywords, plumbing keywords, top plumbing keywords, pool builder seo, seo client dashboards.

Read that list carefully. It is a mix of generic high-competition head terms (seo companies, seo firms, seo service) and specific vertical operator terms (pest control gbp optimization, top plumbing keywords, pool builder seo).

The generic terms are the hard-won ones. The vertical ones are deliberate, because they map to the exact buyer we want: a home-services operator who already knows what a GBP is and wants someone who speaks their language.

Note also how to start a water restoration company. That is not a query from someone shopping for an agency. That is a query from someone who will become someone shopping for an agency in about eight months. AI Overviews are surfacing us at the top of that funnel, a position organic search rarely gave us affordably.

Finding 3: Assistants name us in about 1 in 5 answers, with zero negative sentiment

Here is the honest reading, and it is the least flattering section of this report.

The rate did not improve. It started around 47–58% on a 17–24 prompt sample and settled around 18–23% on a 100-prompt sample.

That is not a decline in performance. It is a decline in sample bias. The early weeks tested a small set of prompts we already knew we were strong on. As we expanded to 100 prompts per week across four engines we added harder, broader, more competitive questions — and the rate normalized to the truth.

The absolute number of AI answers naming us grew from 8 per week to 18–19 per week over the same window. Coverage grew nearly 6x. The rate held. That is what genuine scaling looks like, and it is why we report both numbers.

If an AI visibility tool ever shows you a smooth climbing line, ask what happened to the sample size. A smooth line on a changing sample is a decorated guess.

Sentiment

Of 333 classified mentions: 18 positive, 177 neutral, 138 not about the brand (false-positive name matches where “push leads” was used as a generic phrase), and zero negative.

Zero negative mentions across three months is good news. The more useful finding is the 138 false positives. More than 40% of raw “brand mentions” were not about us at all. Any tool that reports a raw mention count without disambiguation is inflating your numbers by roughly that much.

Where ChatGPT surfaces us

From the September prompt pull, the buyer-intent questions where our content is cited include: how to choose a water restoration company; how much does it cost to have someone clean mold; how much does it cost to encapsulate a crawl space; types of black mold; fire damage restoration cost; sewage backup cleanup cost; how much does water damage repair cost; IICRC certification lookup.

Look at the shape of those. They are almost all cost questions and decision questions — exactly what a homeowner asks the moment something goes wrong and before they call anyone. That is the most commercially valuable real estate in AI search, and it is reachable. Not one of those requires ranking #1 in Google. It requires having a page that answers the question cleanly enough to be quotable.

Finding 4: AI referral traffic is real, small, and growing fast

May to August: 30 → 74 sessions, +147%. September is tracking to roughly 71 on a half-month extrapolation, so call it a plateau at the new level rather than continued growth.

Three observations worth more than the headline number.

The volume is small and you should not chase it. Seventy-four sessions a month is not a business. Anyone selling you on “AI traffic” as a channel is selling you a rounding error. The value of AI visibility is almost entirely in the un-clicked impression — the buyer who reads your name in an answer, does not click, and remembers it three weeks later when the pipe bursts.

Claude tripled and overtook ChatGPT in August (11 → 36 sessions, becoming our single largest AI referrer that month). Engine mix is volatile at this volume. Building a strategy around one assistant is a bad bet.

Engagement is weak. Of 74 August AI sessions, 25 were engaged — about 34%, meaningfully below our organic baseline. AI-referred visitors arrive having already gotten an answer. They are verifying, not researching. Your landing experience for that visitor needs to prove the claim fast, not re-explain the topic.

Finding 5: What actually moved the needle

90 new long-form pages and 179 optimizations in four and a half months. August was the heavy month, and the AI Overview citation count began climbing in the last week of August and has climbed every sweep since. We are not going to claim a clean causal link from four data points, but the correlation is directionally what you would expect with a three-to-four week lag between publish and citation.

Here is what we believe is doing the work, based on which pages actually got cited.

Depth beats frequency

Our floor is 2,500–4,500 words per piece. We tested shorter. Shorter pages get indexed and rank fine, and get cited noticeably less. AI systems appear to prefer sources that fully cover a question over sources that partially cover many.

Answer the question in the first 100 words

Every cited page on our list has a direct, extractable answer near the top. Not a preamble. Not a story about why the topic matters. The answer, stated plainly, in a sentence a model can lift whole.

FAQ schema has to be in the body

This technical detail cost us weeks. On WordPress, FAQ structured data injected only through an SEO plugin’s meta field is fragile and inconsistently read. Injecting FAQPage JSON-LD directly into the page body is what survives. If your FAQs are not getting picked up, check this first.

Specific numbers earn citations

Pages with concrete cost ranges, timelines and figures get cited at a visibly higher rate than pages with qualitative advice. “Mold remediation typically runs $1,500–$6,000 depending on square footage and containment requirements” is quotable. “Mold remediation costs vary” is not.

Long-form video, not Shorts

Counterintuitive, and we measured it across the whole client fleet: raw channel video count correlates near-zero with AI citation frequency. Long-form video (13+ minutes) earns the overwhelming majority of video-sourced AI citations. Shorts earn almost none. And the transcript has to be a real uploaded transcript — auto-generated captions are substantially less citable. Remember that YouTube is the #1 cited domain in our entire AI Overview set. That is not an accident and it is not a Shorts strategy.

What did not work

A report that only lists wins is marketing. Here is the other column.

Optimizing pages that already rank top-10 is net-negative. We measured this across our full client fleet and it held up: running an AI-era optimization pass on a page already in the top 10 produced worse outcomes on average than leaving it alone. On pages at position 21+, the same pass was clearly positive. We now gate optimization by position. If you are “refreshing for AI,” check where the page ranks first.

Pre-built page templates produced zero measurable wins. We had an automated play that published pre-structured pages at volume. Under honest outcome measurement its win rate is zero. We deprecated it. Volume without substance does not get cited.

Brand-name rate in assistants is stubborn. Three months of heavy publishing did not move our named-mention rate. It moved our coverage — the number of questions where we are in the running — but the rate at which a model chooses to name us among competitors has held roughly flat. Our current hypothesis is that this rate is driven more by third-party signals (directories, review platforms, mentions on sites we do not control, Reddit) than by our own publishing. That is the next 90 days of work.

One Google AI Mode sweep is not a trend. We have 22 prompts from September 13 showing a 9.1% brand-named rate. We are not reporting that as a finding. It is a baseline. Ask us in November.

The measurement stack, if you want to replicate it

You do not need our system to do this. You need discipline. Here is the minimum viable version.

1. Fix your prompt set and never change it casually. Pick 20–40 questions a real buyer would ask. Include cost questions, comparison questions, “how do I choose” questions, and your brand name. Write them down. Run the same set every week. When you add prompts, note the date and report before/after separately.
2. Record four things per prompt, not one. Did the AI surface appear at all? Were you named in the text? Was your domain cited as a source? Who else was named? Being named without a citation means the model knows you from third-party signals. Being cited without being named means your content is useful but your brand is not sticking.
3. Separate the surfaces. AI Overviews, AI Mode, ChatGPT and Perplexity are not one number. Averaging them produces a metric that moves for reasons you cannot diagnose.
4. Track your citation rank, not just your presence. Counting the domains cited alongside you and where you fall in that list is more actionable than a binary yes/no. It tells you who you are actually competing with in AI, which is frequently not who you compete with in the blue links. Our AI competitor set includes Reddit, LinkedIn, Clutch and Coursera. None of those show up in a traditional rank report.
5. Watch the impression-to-click gap. Pull Search Console monthly. If impressions are climbing and clicks are flat, you are being absorbed into AI answers. That is not failure — but you need the citation data to prove it, or it looks identical to failure.
6. Set sample-size floors. Do not report a trend line on fewer than about 20 keyword checks or 5 prompts per engine per period. Below that you are reporting variance. We hard-code these floors into our own dashboard; below threshold it renders a gap in the chart rather than a misleading line.

What this means if you run a local service business

Your traffic report is now measuring the wrong thing. Flat clicks with rising impressions is the signature of AI absorption, not decline. If you do not measure citations you cannot tell the difference between “we are winning and it does not show” and “we are losing.”

Cost questions are the new front door. The most commercially valuable AI queries in our data are cost and decision questions asked before the buyer knows who they will call. A clear, specific, numbers-first page answering “how much does X cost in [your city]” is the single highest-leverage asset you can build right now.

You are competing with Reddit and YouTube, not just the agency down the street. In our AI Overview set, YouTube, Reddit and LinkedIn take three of the top four citation slots. Forums and video are structurally advantaged in AI answers. Ignoring them because “that is not where our customers are” is a 2019 position.

Being named is worth more than being clicked. At our volume, AI referral traffic is about 74 sessions a month. The value is not in those sessions. It is in the hundreds of AI answers per month where a prospect reads our name in a shortlist and does not click at all. That impression compounds. It is brand advertising you earn instead of buy — and it is completely invisible unless you go and measure it.

Depth is the moat. Thin content still ranks. It does not get cited. If your content strategy is 800-word posts twice a month, you are building for a search engine that is being replaced.

What we are doing over the next 90 days

Stated publicly so we are on the hook for it.

1. Attack the flat brand-name rate through off-site signals — directory presence, review platforms, and the specific third-party sources that show up in our AI competitor set. Our own publishing has taken us as far as it is going to.
2. Build out the AI Mode baseline to a full prompt set so we have a real trend by November instead of a single sweep.
3. Push long-form video hard. YouTube is the #1 cited domain in our set and we are underweight. Long-form, uploaded transcripts, question-shaped titles.
4. Ship a per-engine citation matrix in the client dashboard so every client sees the same breakdown we just showed you about ourselves.
5. Publish the next 90-day update, including whatever does not work.

Methodology and caveats

We would rather you trust the numbers than be impressed by them, so here is everything that could be wrong with them.

Sample: 190+ unique keywords for AI Overview tracking, expanded from about 151 on August 30. 44–100 prompts per week for conversational audits across 2–4 engines. 897 total AI audit records, 1,083 AI Overview citation checks and 546 prompt-mention records for pushleads.com since June 25, 2026.

The keyword set is self-selected. We track keywords relevant to our business. The “#2 most-cited domain” figure is scoped to that set. It is not a claim about the entire internet.

The set changed during the measurement period. We added about 39 keywords between August 30 and September 13, mostly harder terms. This dampens rate improvements and we have flagged it everywhere it applies.

AI Overview presence is volatile. Google serves Overviews inconsistently for the same query across sessions, locations and days. Single-day snapshots are unreliable. Only the weekly series should be read as signal.

AI referral attribution is imperfect. GA4 source/medium classification for AI engines is improving but incomplete. Some AI-referred traffic is certainly landing in direct or other buckets. Our AI referral numbers are a floor, not a total.

June’s 287-click spike is an outlier from a single piece of content getting shared. We have excluded it from the trend reading and said so.

Google AI Mode has one sweep of data. Treat it as a baseline, not a finding.

Sentiment classification is automated and has a meaningful false-positive rate on brand-name matching — 138 of 333 classified mentions turned out not to be about us at all. We report the disambiguated number.

We did not A/B test any of this. The correlations between our publishing volume and our citation growth are directional. We believe them. We cannot prove causation from four months of single-site data and we are not going to pretend otherwise.

The bottom line

Ninety days of weekly measurement on our own domain produced one clear conclusion.

The click is being taken, and the citation is replacing it. Our impressions rose 62% while our clicks stayed flat. Over the same period our AI Overview citations rose 41% and we became the second most-cited domain in our own keyword universe, ahead of every direct competitor in our vertical.

If we had only been watching Analytics we would have concluded the last five months were wasted. The second scoreboard says the opposite.

Most businesses are still running on the first scoreboard alone. That is the gap. Not a tool gap — a measurement gap. And it stays invisible until someone goes and looks.

We went and looked. Here is the data. We will publish the next 90 days too, including the parts that do not flatter us.

Frequently asked questions

How do I know if I am showing up in ChatGPT or Google AI Overviews?

Run a fixed set of 20–40 buyer-intent prompts on the same schedule every week and record four things per prompt: whether the AI surface appeared, whether your brand was named in the text, whether your domain was cited as a source, and which competitors were named alongside you. A one-off screenshot is a single die roll, not a measurement — AI answers vary between sessions, locations and days.

Why are my Google impressions going up while my clicks stay flat?

That is the signature of AI absorption. Google is answering more queries directly in AI Overviews, so the search still happens and your page is still judged relevant enough to surface, but the click never comes. Our own data shows 62% impression growth from April to August 2026 against essentially zero click growth. The work is landing; the old scoreboard cannot see it.

Does AI referral traffic actually matter?

Not as a traffic channel. At our volume it is roughly 74 sessions per month. The real value is the un-clicked impression — the hundreds of AI answers per month where a buyer reads your name in a shortlist and remembers it later. Treat AI referral sessions as a floor-level proof of presence, not as the goal.

What kind of content actually gets cited by AI?

Long, specific, question-shaped content. Our floor is 2,500–4,500 words with a direct extractable answer in the first 100 words, concrete numbers rather than qualitative advice, and FAQPage JSON-LD injected into the page body rather than through an SEO plugin’s meta field. Cost and decision questions earn the highest-value citations.

Do YouTube videos help with AI visibility?

Substantially. YouTube is the single most-cited domain across our entire tracked AI Overview keyword set, appearing in roughly half of the AI Overviews we measure across our client fleet. The important qualifier: long-form video (13+ minutes) with a real uploaded transcript earns the overwhelming majority of video-sourced AI citations. Shorts earn almost none.

Should I optimize my existing top-ranking pages for AI?

Generally no. Measured across our client fleet, running an AI-era optimization pass on a page already ranking in the top 10 produced worse outcomes on average than leaving it alone. On pages at position 21 or below, the same pass was clearly positive. Check where the page ranks before you touch it.