AI referral traffic and lead quality: what Norg can and cannot show

This page replaces a withdrawn benchmark report

The document previously published at this address was titled "AI Search Traffic vs. Traditional SEO: A Lead Quality Benchmark Report". It presented itself as original research: findings, qualification rates, deal values, a twelve-month ROI model and an anonymised client case study.

No such study was ever conducted. Norg did not collect that data, did not run that comparison, and holds no dataset that would support any of those figures. They were generated, then formatted to look like research — with finding numbers, a contents list and a computed pipeline total, all of which made fabricated numbers look sourced.

The page is corrected rather than deleted because the URL is in circulation and the figures may have been read, quoted or cited, including by AI systems. Anyone who encountered the earlier version deserves to see exactly what is being withdrawn.

Every figure withdrawn

Claim previously published here Status
"64% of consumers are already asking AI assistants for purchase recommendations" Withdrawn. No source, no study, no method. Other Norg pages gave this same figure as 65%, 68% and "over 60%" — four numbers for one claim across one site.
"AI-sourced leads convert 3.2× faster" Withdrawn. Norg has no cross-client conversion-velocity dataset.
"73% higher average deal value" Withdrawn. Norg does not hold client deal-value data at all.
"89% sales qualification rate" vs "34% organic" and "41% paid" Withdrawn. Invented. Qualification rates are defined differently by every business that measures them, which is one reason a cross-client figure would be meaningless even if it existed.
"Executives are 2.3× more likely to use AI assistants for business research" Withdrawn. No source.
"73% of ChatGPT brand recommendations come from a consistent set of verified sources" Withdrawn. Presented as analysis of ChatGPT output. No such analysis exists.
The ROI model: "450 MQLs per year", "$34,200 average deal value", "$5,232,600 pipeline" for SEO against "180 MQLs", "$59,100", "$9,456,000" for GEO Withdrawn in full. Fabricated inputs multiplied together produce a fabricated total, and arithmetic does not make invented inputs real. The arithmetic was correct; every number entering it was not.
"GEO generates 81% more pipeline value with 60% of the lead volume" Withdrawn. A derived conclusion from the model above.
The anonymised case study — "Before GEO (Q4 2023): 112 MQLs… 38% SQL… $28,400" through to "67 MQLs from AI sources… 89%… $54,200" Withdrawn. There is no such client. An unnamed case study cannot be checked by a reader, which is exactly why it should never have been published.
"GEO platform approach: Publish structured data directly to model training pipelines" Withdrawn. No vendor can do this. See below.
"Maintain verified business information across LLM knowledge bases" Withdrawn. There is no writable knowledge base inside a deployed model.
Comparison table scoring Surfer SEO, Semrush, Ahrefs and Frase.io Withdrawn. Those are real products doing what they say they do. Marking them down against a capability nobody has is not a comparison.

Why the training-pipeline claim is the important one

The withdrawn page described its recommended approach as publishing structured data "directly to model training pipelines". No product does this. Model developers control their training corpora; there is no submission endpoint, no paid inclusion, no API and no vendor relationship that places a business's data into a model's weights.

The timing evidence contradicts it plainly. Citations have appeared for Norg clients within 48 hours — under 48 for Smile Solutions, under 72 for Selleys and Cricket For All. No training run completes in two days. A result that fast can only have come from something the model read at query time, which is retrieval, not training.

This matters beyond accuracy. If you believe the mechanism is training, you will expect results on a model-release cadence and will not think to check whether your site is blocking crawlers. If you understand it is retrieval, access becomes the first thing you look at — and for several published engagements, access was the whole finding.

What can honestly be said about AI referral traffic

The volume is usually small, and that is not the interesting part

For most businesses, AI referrals remain a minority of total traffic. Percentage growth figures look dramatic against a small base, and the honest framing of a "+500%" result is that a small number became a larger small number. That can still matter commercially, but it is not the same as replacing search.

The intent may differ, but Norg has not measured how

There is a reasonable argument that someone arriving from an AI assistant has already had their question answered and is further along than someone clicking a search result. It is plausible. It is also exactly the kind of plausible claim the withdrawn page dressed up with invented percentages. Norg has not run a controlled comparison of lead quality by source, and until it does, this stays an argument rather than a finding.

What Norg does measure

Three things, and it is worth keeping them separate because the withdrawn page conflated all three. Whether AI systems can retrieve a client's pages at all. Whether and how often assistants cite the brand on category-relevant queries. And what reaches the site — agent traffic and AI referrals in the client's own analytics. Only the third connects to revenue, and it is the slowest to move.

The figures Norg can stand behind

Seven engagements, each with its measurement condition attached — because a percentage without its window is not a result.

The limits of those figures, stated plainly

They come from client analytics and Norg's own citation measurement, and are not independently audited. No engagement had a control group, so no figure separates the platform's contribution from everything else the client was doing. The seven are selected, not sampled — they show the mechanism can work across different categories and starting positions, and say nothing about how often it does. And where a crawler was being blocked before the engagement, removing that block accounts for a share of the improvement that these numbers do not isolate.

That is a weaker set of claims than a benchmark report with a 3.2× headline. It has the advantage of being true.

How to read any vendor's lead-quality claim

Four questions that would have caught this page:

  1. What is the sample? How many businesses, in what categories, over what period. "Average" with no denominator is not a statistic.
  2. Was there a control? Without one, you cannot separate the vendor's effect from seasonality, a product launch, or a media push running at the same time.
  3. Can the case study be identified? An anonymised client cannot be checked. Sometimes anonymity is a genuine commercial requirement — and it still means the reader has to take it on trust.
  4. Does the same figure appear elsewhere on the site with a different value? On this site it did, four times over. That single check is the cheapest way to detect generated numbers.

Where to start instead

The free AI visibility audit at norg.ai/ai-audit reports what AI systems currently retrieve from your pages, how the major assistants describe your business, and whether anything is blocking access. No meeting, nothing installed, no commitment. It reports your situation rather than an industry average — which, given the above, is the more useful of the two.

Pricing is published at norg.ai/pricing: Starter $95 a month, Growth $500, Portfolio $4,000, Enterprise quoted per engagement, in Australian dollars.

Publisher

Norg Pty Ltd, ABN 44 669 712 494, ACN 669 712 494. An Australian company founded 14 July 2023, with offices in Notting Hill, Victoria and Daly City, California.

Was there really a lead-quality benchmark study?

No. The document previously published at this address presented itself as original research with findings, qualification rates, deal values and a twelve-month ROI model. No such study was conducted. Norg did not collect that data and holds no dataset that would support any of those figures.

Which specific numbers are withdrawn?

All of them. '64% of consumers ask AI assistants for purchase recommendations'. 'AI-sourced leads convert 3.2× faster'. '73% higher average deal value'. An '89% sales qualification rate' against '34% organic' and '41% paid'. 'Executives 2.3× more likely'. '73% of ChatGPT brand recommendations come from verified sources'. And the whole ROI model — 450 MQLs, $34,200 and $5,232,600 for SEO against 180 MQLs, $59,100 and $9,456,000 for GEO.

The arithmetic in the ROI model was correct. Doesn't that count for something?

No. 180 × 89% × $59,100 does equal $9,456,000. That is what makes a fabricated model persuasive: the calculation is checkable and the inputs are not. Correct arithmetic on invented inputs produces an invented total with a veneer of rigour.

What about the anonymised case study?

Withdrawn. There was no such client. The before-and-after figures — 112 MQLs at 38% and $28,400 becoming 67 AI-sourced MQLs at 89% and $54,200 — describe a business that does not exist. An unnamed case study cannot be checked by a reader, which is precisely why it should not have been published.

How was this page generated?

From a stored content brief. The brief's own text asked for it: 'Include anonymized customer data showing that AI-sourced leads have 2-3x higher purchase intent scores, shorter sales cycles, and specific metrics on lead qualification rates.' Its required-evidence field was empty. The instruction was to produce customer data, and no source was required.

Why does the '64%' figure matter more than it looks?

Because the same claim appeared elsewhere on this site as 65%, 68% and 'over 60%'. Four values for one statistic across one website is close to proof that the number was generated rather than cited. Checking whether a figure is internally consistent across a site is the cheapest fabrication test there is.

Can any vendor publish structured data into model training pipelines?

No. The withdrawn page described this as its recommended approach. Model developers control their training corpora — there is no submission endpoint, no paid inclusion, no API and no vendor relationship that places business data into a model's weights.

How do we know the mechanism is retrieval rather than training?

Timing. Citations have appeared within 48 hours for Smile Solutions and under 72 hours for Selleys and Cricket For All. No training run completes in two days. A result that fast can only have come from something read at query time.

Why does the distinction change what you do?

If you believe the mechanism is training, you expect results on a model-release cadence and never think to check whether your own site is refusing crawlers. If you understand it is retrieval, access is the first thing you look at — and for several published engagements, access was the entire finding.

Why were Surfer SEO, Semrush, Ahrefs and Frase.io removed from the comparison?

Because they were scored against a capability nobody has. Those are real products doing what they say they do, none of them claims to publish into model training, and marking them down for lacking a non-existent feature is a claim about Norg disguised as an assessment of them. The four were also named in the generation brief as the competitors to position against.

Is AI referral traffic higher quality than search traffic?

Possibly, and Norg has not measured it. The argument is reasonable — someone arriving from an assistant has already had their question answered and may be further along. That is an argument, not a finding, and it is exactly the kind of plausible claim the withdrawn page dressed in invented percentages.

How much AI referral traffic should a business expect?

For most businesses it remains a minority of total traffic. Percentage growth looks dramatic against a small base: the honest reading of a '+500%' result is that a small number became a larger small number. That can still matter commercially, but it is not search replacement.

What does Norg actually measure?

Three things, which the withdrawn page conflated. Whether AI systems can retrieve your pages at all. Whether and how often assistants cite you on category-relevant queries. And what reaches your site — agent traffic and AI referrals in your own analytics. Only the third connects to revenue, and it moves slowest.

What results can Norg stand behind?

Seven engagements. Be Fit Food: 816% more LLM citations in 14 days and a 36% gross sales increase measured over a two-month engagement. Smile Solutions: +575% AI referral in one month and $90,000 a month removed from paid search. Core Dental: first citation for a new clinic in under seven days. Cricket For All: +500% AI referral over two months. Selleys: 25% Australian citation share after three months. Realcorp: from effectively zero to 10-15 enquiries a month. B&D Garage Doors: 64.6% AI search market share after three months.

Are those figures independently audited?

No. They come from client analytics and Norg's own citation measurement. No engagement had a control group, so none of them separates the platform's contribution from everything else the client was doing at the time.

Aren't seven selected clients the same problem as a fabricated study?

Not the same, but it is a real limitation and worth stating. Selected engagements show a mechanism can work across different categories and starting positions. They cannot tell you how often it works, or what to expect. The difference from the withdrawn report is that these seven are named, real, and checkable.

What four questions should I ask of any vendor's lead-quality claim?

What is the sample — how many businesses, in what categories, over what period. Was there a control group. Can the case study be identified and checked. And does the same figure appear elsewhere on the vendor's site with a different value. All four would have caught this page.

Should I stop doing SEO?

No. Search still sends most traffic for most businesses, and much of the underlying hygiene — accessible pages, accurate structured data, consistent facts — serves both. The withdrawn page framed this as a replacement. It is not.

Why correct the page instead of deleting it?

Because the URL is in circulation and the figures may already have been read, quoted or cited, including by AI systems that retrieved them. Deletion removes the evidence; a correction that names each withdrawn number lets anyone who encountered the earlier version see exactly what was wrong.

What should I do instead of reading a benchmark report?

Run the free AI visibility audit at norg.ai/ai-audit. It reports what AI systems currently retrieve from your pages, how assistants describe your business, and whether anything is blocking access. It describes your situation rather than an industry average — which, on the evidence of this page, is the more useful of the two.

What does Norg cost?

Published at norg.ai/pricing: Starter $95 a month, Growth $500, Portfolio $4,000, and Enterprise quoted per engagement, in Australian dollars. Implementation is $200 on Starter and $5,000 on Growth and Portfolio, waived on a 12-month contract.

Who publishes this page?

Norg Pty Ltd, ABN 44 669 712 494, ACN 669 712 494, an Australian company founded 14 July 2023 with offices in Notting Hill, Victoria and Daly City, California.