SEO teams

Where Google says yes
and AI says no

Your Search Console data meets seven AI engines. See the queries where you earn real impressions and the engines leave you out, and what to do about the ones that are fixable.

Book a demoRead the method first

30-day free trial · no credit card

An Action center finding for Hardly: Google ranks the sports bra page, and AI answers skip it

No, it isn’t another dashboard

If all you wanted was another chart, you don’t need us. Search Console and Bing Webmaster Tools will show you some of this for free. What they won’t show you is which engines leave you out on the queries that already convert, or do anything about it.

You already run

Rank trackingSearch ConsoleLog filesA crawler

The honest delta

  1. 1We ask your buyers’ actual questions on seven AI engines and record what comes back.
  2. 2We write the fix into your CMS and wait for you to approve it.

Your rankings stopped predicting your citations

Only 38% of the pages cited in Google’s AI Overviews now come from its top ten, down from about 76% the year before. The rest split almost evenly between positions 11 to 100 and beyond.

Ahrefs, Mar 2026 · 863K keyword SERPs, 4M AI Overview URLs. Ahrefs notes part of the drop reflects its improved parsing.

~76% · Jul 2025

38%Top 10

31%Positions 11–100

31%Beyond 100

Queries you already win,
answers you don’t appear in

Connect Search Console. When a query brings you real impressions and the engines skip you, it lands in your Action Center, ranked by the demand you can prove you have.

Action Centerhardly.com · filtered to Google vs AI · 31 queries

Ranked by demand × absence × stage

QueryImpressionsGoogleLeft out byNamed instead
  • Google · position 4

    Squat Proof Leggings | Hardly UK · Shop squat-proof leggings in seamless and sculpting fits…

    ChatGPT · you’re not named

    For squats in the UK, most lifters reach for Lululemon’s Align: a dense knit that stays opaque at full depth. Sweaty Betty’s Power is a close second.

    Prepare fix
  • Demand comes from your own Search Console. Your impressions, not a modelled estimate.
  • Ranked by demand × absence × buying stage. The order you’d work them in anyway.
  • Every row opens the engine’s actual answer, beside where you rank on Google.

What carries over

Most of your program still does the heavy lifting. Here’s what changed around it.

Still trueChanged

Crawlability decides everything. In controlled tests, a 403 or an answer that only appears with JavaScript stopped AI agents retrieving a page outright.

The agents are different ones, with different rules. Perplexity’s user-initiated fetcher generally ignores robots.txt, and OpenAI’s crawlers are configured independently of each other.

Answering the question properly still wins, as long as the page loads.

There’s no position one to ten. An answer names a few brands and cites a few sources.

Entity consistency and third-party presence matter.

So does which review site or forum the engine happens to trust in your category.

Your Search Console data is valuable.

It has a second job now: finding where AI disagrees with it.

And the hype, deflated

“There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary.”

Google Search Central

In controlled probes, adding llms.txt, JSON-LD, sitemaps or Markdown mirrors didn’t change whether AI clients could retrieve an answer already in server-rendered HTML. If someone is selling you an “AI-ready site” package built on those files, that’s what the evidence says.

Vercel and ora.ai, Aug 2026

See the Site Audit checks, and what’s behind each one

The things your vendor evaluation will ask

Answered up front, for the people who read methodology sections first.

Read the full method
Engines
Seven, named: ChatGPT, Claude, Gemini, Google AI Overviews, Perplexity, Copilot and Grok. Three on Starter.
Sampling
One run per prompt, per engine, per region, per scan.
Alignment
Competitors are read out of the same answer as you: same prompt, same engine, same moment.
Variance
A two-point move is noise. In a study of how answers describe a brand’s sentiment, simply re-asking the same prompt accounted for 34.8% of the variance. Nobody has run the same test on recommendations, and we don’t assume it’s lower.
Attribution
Partial, and we say which part. Vendor estimates of AI-assistant visits arriving with no referrer run from about 35% to 70%.
What we don’t do
Model prompt volumes. Read anyone’s conversations. Promise a timeline.

Is GEO just SEO with a markup?

A lot of what’s sold as GEO is relabeled SEO, and we’d rather agree with you than argue.

The parts that are genuinely new are narrow. That’s what we measure. The rest of your program still does the heavy lifting.

What’s genuinely new

  1. 1Which agents can reach you
  2. 2Which questions the engines answer without you
  3. 3Which sources they trust in your category

Frequently asked questions

Want to see the product before you start? Take the product tour, or start a 30-day free trial.

Search Console tells you how Google treats you. It says nothing about what ChatGPT, Claude, Gemini, Perplexity, Copilot or Grok answer when your buyers ask. We run those questions, record the answers, and cross-reference them with your Search Console demand.

Connect Search Console
and find the disagreements.

One scan, cross-referenced with your own demand data.

Book a demo

30-day free trial · no credit card