Does schema markup get you cited by AI?

Mostly no. Ahrefs ran a causal test on 1,885 pages in May 2026 and found JSON-LD schema made no significant difference to AI citations. FAQ schema is the exception, carrying around 40% of ChatGPT source-selection weight. Keep schema for rich results and entity clarity, but do not expect product or Organization markup to earn a citation.

Is llms.txt worth adding?

Not as a citation tactic in 2026. Only about 10% of scanned domains have one, and no major engine shows a measured citation lift from it. A 90-day client test moved Claude citations by under one percentage point. Ship it as cheap future-proofing for agentic AI, but do not let it displace proven work.

What is wrong with AI monitoring tools?

Nothing, as long as you treat them as a thermometer, not a cure. A dashboard that tells you that you are invisible is a diagnosis. It does not change a single citation. The work that follows, fixing extractability, freshness and authority, is what moves the number, and most monitoring tools do none of it.

What is the difference between an AI mention and an AI citation?

A mention is the model naming your brand from training data, with no live link. A citation is the engine pulling your page as a source and linking it in the answer. Mentions are nice but unverifiable and unstable. Citations send traffic and authority. Count citations.

Can any agency guarantee AI citations?

Be sceptical of anyone who does. Citations are decided by third-party engines no agency controls, so guaranteeing a specific citation is overpromising. What can be guaranteed is the input work. We guarantee every link we place stays live for 12 months or we replace it. Demand guarantees on the inputs, not the output.

What should I actually pay for in GEO?

Three things that clear the citation floors: extractable content, freshness, and off-site authority. Plus honest measurement of citation share across the engines. If a line item does not raise one of those floors or measure the result, question it. Most wasted GEO spend goes on tactics that do neither.

GEO That Doesn't Work: What To Stop Paying For

By Matt, Co-founder, GeoLinks · 23 June 2026 · 4 min read

A marketing budget sheet with several AI-search line items crossed out in red, editorial flat-lay style in natural light

Most of what is sold as AI-search optimisation does nothing measurable. Schema bolted onto every page, llms.txt files, monitoring tools that only report bad news, and vanity “mentions” all fail to earn a citation. This post names the four tactics to stop paying for, the work that actually moves citations, and the honest truth about citation guarantees. The data is drawn from independent 2026 studies and our own client tests.

We run a GEO business, so naming the dead tactics costs us upsells. We do it anyway, because the alternative is billing for theatre and losing the client when it does not work.

The four things you can stop paying for

GEO is young enough that the market is full of confident tactics with no evidence behind them. Four come up again and again. Each one feels like progress. None of them earns a citation on its own.

Schema theatre

Adding JSON-LD to every page is the most common false win in GEO. It feels technical and complete. It does almost nothing for citations.

Ahrefs tested this directly across 1,885 pages in May 2026. Pages carrying schema were cited at the same rate as pages without it. We add schema to every site we run, including this one, for rich results and entity clarity. We do not pretend it earns citations. FAQ schema is the one real exception, and the full data is in the schema markup study.

The llms.txt distraction

The llms.txt file is sold as the way to “talk to AI”. The adoption numbers tell the real story. Only about 10% of scanned domains have shipped one.

More to the point, no major engine shows a measured citation lift from having one. A 90-day client test moved Claude citations by under a single percentage point. It is cheap future-proofing for agentic AI in 2027 and beyond, not a 2026 citation lever. The honest data is in the llms.txt breakdown.

Monitoring-only tools

A wave of tools will now tell you how visible you are in AI search. That is genuinely useful as a diagnosis. The problem is when the dashboard is the whole product.

A tool that reports you are invisible has not changed anything. It is a thermometer, not a treatment. The fix is the work that follows: rewriting pages to be extractable, holding freshness, and building off-site authority. Pay for the diagnosis if it is good. Do not mistake it for the cure.

Vanity mentions

The last false win is the “mention”. A model names your brand in an answer, with no link, and a tool counts it as a win. It feels like visibility.

A mention pulled from training data is unverifiable and unstable. It sends no traffic and it can vanish on the next model update. A citation is different: the engine pulls your live page as a source and links it. Traffic and authority follow the citation, not the mention.

Theatre versus fix

Tactic	What it feels like	What it does	Verdict
Schema on every page	Technical completeness	No citation lift (FAQ schema aside)	Keep for SEO, not GEO
llms.txt file	Talking to AI	No measured citation lift	Future-proofing only
Monitoring dashboard	Being in control	Reports, does not fix	Diagnosis, not cure
Counting mentions	Visibility	Unverifiable, unstable	Count citations instead
Extractable pages	Slow, unglamorous	Directly earns citations	Pay for this
Off-site authority	Expensive	Clears the trust floor	Pay for this

What actually fixes it

The fixes are less exciting than the theatre, which is exactly why they get skipped. Three floors decide whether a page gets cited: it has to be extractable, it has to be fresh, and the brand has to hold enough off-site authority. Clear all three and you become eligible to be cited. Stay below any one and you stay invisible.

That is the Citation Floor Method in one line. The work is remediation, not decoration: rewrite the opening of each commercial page into a 50 to 75 word answer block, hold a real 30-day refresh, and earn third-party corroboration, because around 91% of citations come from sources you do not own. Then measure. The five numbers that prove it is working are in the AI citation reporting template.

The citation guarantee gap

Here is the part almost nobody in GEO will say out loud. No agency can honestly guarantee a specific AI citation, because citations are decided by third-party engines that no provider controls.

That is why the “citation guarantee” is an empty space in the market. The honest version is to guarantee the inputs, not the output. We guarantee every link we place stays live for 12 months or we replace it free, set out in writing on our guarantee page. We will not promise a named citation by a named date, because anyone who does is selling certainty they cannot deliver.

So when you compare providers, invert the usual test. Be most sceptical of the one promising guaranteed citations. Trust the one guaranteeing the work that earns them, and showing you the citation share before and after.

Is GEO just SEO? The honest answer sets out where the two overlap and where GEO genuinely adds work.
Why isn’t my brand in ChatGPT? Seven real reasons is the self-diagnosis to run before paying anyone for anything.

Want the diagnosis without the theatre? The free AI Visibility Check scans all five engines and shows your real citation share in under five minutes. When you want the fixes done, our pricing lists every tier with no gated quote.