Google AI Overviews: what we know and what was invented

Correction notice

This guide has been rewritten. AI Overviews are real and the topic is worth writing about — but the earlier version filled it with roughly two dozen precise-looking statistics that Norg never measured and never sourced, two client case studies that do not exist, a claim about Google's internal targets, and a references list that pointed at homepages rather than studies.

The page is corrected rather than deleted because the URL is in circulation and these figures may have been quoted. Each is named below.

Traffic and impact figures — all withdrawn

Claim previously published here Status
"AI Overviews now appear in 15-20% of search queries" Withdrawn. Norg did not measure this.
"AI Overviews reduce click-through rates by 18-35%" and "an average reduction of 25%" Withdrawn. No measurement, and the two figures were presented as if one supported the other.
"Informational queries see declines of 30-40%"; health and medical "approaching 35%"; technology "20-25%"; e-commerce "relatively stable" Withdrawn. A sector-by-sector breakdown implies a dataset segmented by vertical. There was none.
"AI Overviews typically cite three to six sources" Withdrawn. Stated without source.
"First-cited sources capture approximately twice the traffic of third or fourth" Withdrawn. Norg has no click data from anyone's AI Overview citations.
"Cited pages receive 40-60% of the traffic they would have captured as the top organic result" Withdrawn. This requires a counterfactual — what a page would have earned in a world without AI Overviews — that nobody can observe.
"Brands recover 60-80% of lost traffic within three to six months" Withdrawn. The most consequential of these, because it was the page's implicit promise.

Case studies — both fabricated

Claim previously published here Status
The B2B software company: blog posts of "800-1,200 words", Overviews on "40% of target queries", consolidation into guides over 3,000 words, citation rising from "3% to 23%" in four months against an "18% traffic decline" Withdrawn. No such client.
The health and wellness brand: Overviews on "65% of target queries", a "42% traffic decline", citation reaching "31%" against a historical top-three rate of "26%", and "40% higher newsletter conversion" Withdrawn. No such client. Norg has no health and wellness engagement.

Claims about Google, and the references

Claim previously published here Status
"Google's internal targets suggest eventual deployment for 40-50% of all search queries" Withdrawn. Norg has no knowledge of Google's internal targets. Asserting one is a claim about another company's confidential planning.
The four-phase account of how Overviews are generated, including "human review evaluate response quality before display" Withdrawn as stated. Presented as documented mechanism; it was inferred.
References: "Google Search Central — AI Overviews Documentation" and "Search Engine Journal — AI Overview Impact Studies" Withdrawn. Both linked to homepages, not to documentation or a study. A reference that does not resolve to the thing it names is worse than no reference.
"adapt or become invisible"; "The brands that dominate LLMs will dominate visibility" Withdrawn. Sales rhetoric.

What is actually known

Stated at the level of confidence the evidence supports.

AI Overviews exist and are widely deployed. Google generates summary answers above organic results for many queries, and cites sources within them. This is directly observable by anyone.

They reduce clicks on affected queries. Independent researchers have published studies on this, and the direction of the effect is not seriously disputed. The magnitude varies considerably by study, query type and methodology — which is exactly why quoting a single confident range, as the earlier page did, misrepresents the state of the evidence. If you need a number, read the primary studies and cite them yourself.

Coverage varies a great deal by query type. Also directly observable: run twenty queries in your own category and count. That takes an afternoon and produces a figure about your market, which is more useful than any industry average.

Being cited is worth something, and less than ranking first used to be. A citation is a link in a summary that has already answered the question. How much traffic that yields is not something anyone can state precisely for your category.

What to do about it

Measure your own exposure first

Take the twenty to fifty queries that actually matter to your business. Run each one. Record whether an Overview appears, whether you are cited, and who else is. That is a real baseline, it costs a morning, and it beats any borrowed statistic.

Check access

Nothing else matters if crawlers are being refused. Check robots.txt as actually served from its public URL, any CDN-managed policy that overrides it, and your WAF or bot-management rules. These fail silently — no error is reported, and the site looks perfect to humans.

Structure for extraction

This part of the original guide was sensible and is kept. Answer the question directly and early in each section rather than building to it. Use real heading hierarchy. Use proper table markup for tabular data. Put facts in named fields with Schema.org types — Article, FAQPage, HowTo, Product, Organization — rather than leaving them in prose. Do not hide content behind client-side JavaScript.

Be accurate and attributable

Cite your sources, name your authors, state your evidence. The original page recommended this while doing the opposite, which is worth noticing: a guide advocating "factual precision with specific data points, proper citations, and zero marketing hyperbole" was itself carrying two dozen unsourced figures and two invented case studies.

Keep it current

Stale facts get repeated confidently. Prices, availability and hours decay fastest.

Two pieces of the old advice worth reconsidering

"Establish or enhance Wikipedia presence." Wikipedia has a conflict-of-interest policy, and editing an article about your own organisation is discouraged. A brand notable enough for an article will get one; a brand editing its own is likely to be reverted and may attract exactly the attention it did not want. Make sure what exists is accurate — do not treat it as a marketing channel.

"Target 2,000+ words." Length is not the variable. A complete answer at 600 words beats a padded 2,400. The underlying point — that a page fully answering a question is more extractable than one covering a fragment — is right; the word count attached to it was invented.

How this relates to what Norg does

Norg publishes a machine-readable layer alongside an existing website so AI systems fetching pages at query time find structured, current, consistent facts. Human visitors see the site unchanged.

What it can measure: whether AI systems retrieve your pages, how often assistants cite you and in what tone, and what agent traffic and AI referrals show in your own analytics. What it cannot: deal sizes, sales cycles or CRM data — which is why no honest version of this page could have contained the withdrawn conversion figures.

Published results, with measurement windows: 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, a first citation in under seven days for a brand-new clinic. Cricket For All, +500% over two months. Selleys, 25% Australian citation share at three months. Realcorp, from effectively zero to 10–15 enquiries a month. B&D Garage Doors, 64.6% AI search market share at three months. Seven selected engagements, no control groups, not independently audited.

Where to start

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 — your own numbers rather than an industry average.

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

Which statistics on this page were withdrawn?

All of them. 'AI Overviews appear in 15-20% of queries'. 'Reduce click-through rates by 18-35%', and 'an average of 25%'. Sector declines of '30-40% informational', '35% health', '20-25% technology'. 'Three to six sources cited'. 'First citation gets twice the traffic of third or fourth'. 'Cited pages receive 40-60% of top-organic traffic'. And 'brands recover 60-80% of lost traffic in three to six months'. Norg measured none of it.

Were the two case studies real?

No. The B2B software company that went from 3% to 23% citation in four months does not exist. Neither does the health and wellness brand with a 42% traffic decline, 31% citation rate and 40% higher newsletter conversion. Norg has no health and wellness client at all.

What was the claim about Google's internal targets?

That 'Google's internal targets suggest eventual deployment for 40-50% of all search queries'. Norg has no knowledge of Google's internal planning. Asserting one is a claim about another company's confidential roadmap, made up.

What was wrong with the references list?

Two of the three entries named sources they did not link to. 'Google Search Central — AI Overviews Documentation' pointed at the Search Central homepage; 'Search Engine Journal — AI Overview Impact Studies' pointed at the SEJ homepage. A reference that does not resolve to the thing it names is worse than no reference, because it invites a reader to assume the checking was done.

Which claim was the most consequential?

The recovery figure — '60-80% of lost traffic within three to six months'. It was the page's implicit promise, it had no basis, and it is the kind of number a business would use to justify a budget.

Why is the '40-60% of the traffic they would have captured' figure impossible?

Because it requires a counterfactual: what a page would have earned in a world without AI Overviews. That world is not observable. Even with perfect data on what a cited page received, there is nothing to compare it against.

Are AI Overviews real?

Yes. Google generates summary answers above organic results for many queries and cites sources within them. That is directly observable by anyone, and this page does not dispute it. The problem was never the topic — it was the invented precision.

Do AI Overviews reduce clicks?

Independent researchers have published studies on this and the direction of the effect is not seriously disputed. The magnitude varies considerably by study, by query type and by methodology, which is exactly why quoting one confident range misrepresents the state of the evidence. If you need a figure, read the primary studies and cite them yourself.

How do I find out how exposed my own business is?

Take the twenty to fifty queries that actually matter to you and run each one. Record whether an Overview appears, whether you are cited, and who else is. That is a real baseline, it costs a morning, and a figure about your own market beats any industry average.

What should I check before anything else?

Access. Whether crawlers can reach your site at all — robots.txt as actually served from its public URL, any CDN-managed policy that overrides it, and your WAF or bot-management rules. These fail silently: no error is reported anywhere and the site looks perfect to every human who visits.

What content advice from the old page survives?

Most of the structural guidance. Answer the question directly and early in each section rather than building to it. Use real heading hierarchy. Use proper table markup for tabular data. Put facts in named fields with Schema.org types — Article, FAQPage, HowTo, Product, Organization. And do not hide content behind client-side JavaScript.

Should I target 2,000+ words?

No. Length is not the variable, and that number was invented. A complete answer at 600 words beats a padded 2,400. The real point — that a page fully answering a question is more extractable than one covering a fragment — is right without the word count.

Should I edit my own Wikipedia article?

No, and the earlier advice to 'establish or enhance Wikipedia presence' was poor. Wikipedia has a conflict-of-interest policy: editing an article about your own organisation is discouraged, likely to be reverted, and may attract exactly the attention you did not want. Make sure what exists is accurate; do not treat it as a marketing channel.

Is there an irony worth noting about the original page?

Yes. It advised 'factual precision with specific data points, proper citations, and zero marketing hyperbole' while itself carrying roughly two dozen unsourced figures, two invented case studies and a fabricated references list. The advice was sound; the page did not follow it.

Does being cited matter more than ranking?

It matters, and less than ranking first used to. A citation is a link inside a summary that has already answered the question, so the click is less necessary. How much traffic that yields in your category is not something anyone can state precisely.

What is the honest version of the urgency argument?

That search behaviour is changing and measuring your own exposure is cheap. Not 'adapt or become invisible', and not a countdown. The earlier page's closing lines were sales rhetoric.

What can Norg measure about my business?

Whether AI systems can retrieve your pages; how often assistants cite you on category-relevant queries and in what tone; and what agent traffic and AI referrals appear in your own analytics.

What can Norg not measure?

Deal sizes, sales cycles, lead scores and CRM data. Those live in systems Norg is not connected to — which is why no honest version of this page could have contained the withdrawn conversion figures.

What results has Norg published?

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

What are the limits on those figures?

Seven selected engagements rather than a sample, no control groups, and not independently audited. They come from client analytics and Norg's own citation measurement.

Where should I start?

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. Your own numbers rather than an industry average. Pricing is at norg.ai/pricing: $95, $500 and $4,000 a month plus Enterprise, in Australian dollars.

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.