In mid-2025, we modeled a progressive deterioration of search-dependent web performance driven by AI Search in two structural changes: increasing zero-click behavior within search engines and the growing tendency of users to obtain answers directly from LLMs rather than initiating a traditional Google or Bing search.

The original downside model assumed approximately −15% after 12 months, −25% after 24 months and −35% after 36 months across traffic, ecommerce revenue or lead generation depending on the property.

One year later, the disruption is clearly real, but its impact is not uniform and cannot be represented accurately by a single decline coefficient.

The emerging conclusion is:

AI-driven disruption materialized, but exposure is highly dependent on business model and query intent. Transactional and relationship-driven properties have so far proven resilient, while informational publishing shows much stronger evidence of absolute traffic substitution.

The two properties where a dedicated LLMs OSE / GEO optimization program was deployed were not only protected from the modeled decline; they showed improving search and business performance. This remains observational evidence rather than experimental proof of causality.

Website types analyzed: Transactional Ecommerce, B2B Lead Generation, Secondary-Market Ecommerce, Editorial Publisher

Analysis: 2025 forecast vs. 2026 observed evidence

Executive Comparison

Site 1	Transactional Ecommerce	Yes — multiple intervention waves	Organic sessions +34.7%; Organic revenue +75.9%; Total revenue +105.8% projected YoY	Clicks +11.2% despite impressions −34.4%	Strongly outperformed downside scenario
Site 2	Corporate / B2B Lead Gen	Yes — intervention began after forecast	Total traffic remains above modeled downside	Clicks +12.3%; impressions +15.2%; ranking strongly improved	Forecasted search decline not observed
Site 3	Transactional Ecommerce / secondary market	No	Organic sessions −17%; organic revenue −29.7% projected YoY	Clicks −21.8%, impressions −34.8%, while CTR and ranking improved	Organic weakness real, but not caused by ranking deterioration
Site 4	Editorial / Informational Publisher	No	Organic sessions −53.5%	Remaining audience substantially more engaged	Strongest evidence of informational-search substitution

1. The Original Forecast

The original model was intentionally broader than Google AI Overviews.

It identified three related structural risks.

First, zero-click search. The 2025 analysis noted that:

“~60% of Google searches already result in no clicks.”

Second, growing AI-generated answers directly inside search engines.

Third, and most importantly for this retrospective:

“Platforms like ChatGPT, Perplexity, and Gemini now intercept a growing share of informational and navigational queries.”

That means disruption can happen before a user ever reaches Google, not merely after an AI Overview appears in a SERP.

The operational forecast simplified these forces into one stress model:

Horizon	Assumed reduction in search-driven performance
12 months	−15%
24 months	−25%
36 months	−35%

The same proportional loss was then applied to traffic, search-attributable ecommerce revenue, or leads depending on the business model.

As the original methodology explicitly stated, the objective was a model that was:

“intentionally conservative and transparent”

and designed for stakeholder scenario planning before more sophisticated behavioral models became possible.

One year of observed data now allows that model to be refined.

2. LLM Disruption Happens at Three Different Layers

A major lesson from the retrospective is that standard analytics captures only part of the change.

Layer 1 — Demand displaced before Search

A query that would historically have started on Google may now start directly inside ChatGPT, Gemini, Claude, Perplexity or another assistant.

If the user receives a satisfactory answer without clicking:

Google Search Console records nothing. GA4 records nothing.

This is particularly important for informational demand.

A decline in Google impressions may therefore represent not only changing rankings or conventional market demand, but potentially a reduction in the number of queries that ever enter a traditional search engine.

Layer 2 — Cannibalization inside Search

The user still chooses Google or Bing, but AI-generated SERP experiences satisfy some or all of the informational need before the click.

This part is partially observable through:

Impressions → Ranking → CTR → Clicks

Layer 3 — Traffic recovered from AI systems

Some AI interactions still generate website visits.

Those referrals can increasingly be measured in analytics by grouping traffic from ChatGPT, Gemini, Perplexity and similar platforms.

But these referrals represent only the clickable residue of AI consumption.

They do not measure users who received the information, considered the brand, and never visited the source website.

3. Site 1 — Transactional Ecommerce Proved Highly Resilient

Site 1 received several dedicated LLM-oriented Search Optimization interventions between 2025 and 2026.

Its commercial trajectory became the strongest counterexample to the original proportional-loss assumption.

KPI	Projected 2026 YoY
Total Sessions	+49.1%
Organic Sessions	+34.7%
Paid Sessions	+16.7%
Total Ecommerce Revenue	+105.8%
Organic Revenue	+75.9%
Paid Revenue	+45.0%
Organic + Paid Revenue	+60.1%

Even before annualizing the year, the site had already approximately matched the entire previous year’s traffic, while total ecommerce revenue was already roughly 38% above the full previous year.

Organic revenue was already approximately 18% above the previous full-year result.

Most importantly, the current search-revenue run rate sits approximately 143% above the revenue level implied by the original downside model.

Search behavior

The Google Search data provide an important explanation:

Metric	YoY change
Clicks	+11.2%
Impressions	−34.4%
CTR	+69.6%
Average ranking	Strong improvement

The website therefore generated more clicks from substantially fewer impressions.

The search footprint became more efficient and more concentrated around stronger positions and/or higher-value queries.

This demonstrates why a decline in overall Search visibility cannot automatically be translated into an equivalent loss of commercial value.

AI-originated acquisition

AI referrals also became measurable.

By the end of the observed 2026 period:

• AI-attributed sessions were approximately +146% versus the entire previous year.
• AI-attributed ecommerce revenue was approximately +249% versus the entire previous year.

AI therefore appears simultaneously as a cannibalization risk and an emerging acquisition channel.

4. Site 2 — B2B Lead Generation Remained Search-Resilient

Site 2 operates in a significantly narrower, more technical and more B2B-oriented search environment.

A dedicated LLM Search Optimization intervention began after the original forecast.

The subsequent Search comparison shows:

Metric	YoY change
Google Search Clicks	+12.3%
Impressions	+15.2%
CTR	−2.5%
Average ranking	19.3 → 7.7

The small CTR compression is compatible with a more answer-heavy SERP environment, although it cannot be attributed specifically to AI.

What matters is the net outcome:

visibility increased sufficiently to generate more absolute traffic despite slightly lower click propensity.

The projected total traffic level is approximately 18% above the original downside target.

This produces an important lesson:

CTR compression and traffic growth can occur simultaneously.

The site’s original lead-loss projection cannot be audited on a like-for-like basis because the conversion measurement definition changed during the observation period. For methodological integrity, no artificial lead comparison was produced.

5. Site 3 — Organic Weakness Without SEO Ranking Weakness

Site 3 operates an ecommerce model in a smaller geographic market and received no dedicated GEO/OSE intervention.

Here, the organic decline anticipated in the original risk model is much more visible:

KPI	Projected YoY
Total Sessions	+4.8%
Organic Sessions	−17.0%
Paid Sessions	+8.3%
Total Revenue	−2.3%
Organic Revenue	−29.7%
Paid Revenue	+12.1%
Organic + Paid Revenue	+7.7%

The organic revenue component alone performs approximately in line with — and slightly worse than — the original downside expectation.

Yet Search Console reveals a very different mechanism from conventional SEO deterioration:

Metric	YoY change
Search Clicks	−21.8%
Impressions	−34.8%
CTR	+20.0%
Average ranking	Improved

The site is ranking better and earning a higher CTR whenever it appears.

What contracted is the search opportunity itself.

Possible explanations include lower market demand, a smaller query footprint, product-demand changes and the upstream displacement of some informational searches toward LLMs.

The available data cannot causally separate these factors.

The result nevertheless highlights a crucial point:

LLM-era search decline does not need to look like an SEO ranking decline.

A website can improve its competitiveness inside Google while the pool of traditional searches available to it becomes smaller.

6. Site 4 — The Strongest Evidence of Informational Search Compression

Site 4 is an editorial publisher.

This makes it structurally different from the other websites because, for many queries, the information itself is the product.

The analysis compared two consecutive annual periods covering the same language and content market before and after a domain migration.

The migration remains an important confounding variable, but the new dataset was already filtered to the equivalent language section, making the pre/post comparison substantially more robust.

KPI	Change
Total Sessions	−53.1%
Organic Search Sessions	−53.5%
Organic Engaged Sessions	−49.8%
Organic Engagement Rate	+5.3 percentage points
Avg. Engagement Time / Organic Session	+13.4%
Organic Events / Session	+34.0%

The original model expected approximately −15% traffic after one year.

The observed decline was more than 3.5 times larger.

This is the one property where the original concern not only materialized, but substantially exceeded the stress scenario.

Less traffic, better traffic

The most interesting signal is not the decline alone.

As Organic Search traffic contracted by more than half:

• engagement rate increased;
• average engagement time increased;
• events per session increased substantially.

This pattern is highly consistent with informational traffic pruning.

Historically, publishers receive large quantities of low-commitment search traffic from relatively simple informational questions.

Today, some of those needs can be fulfilled either:

1. before Search, by asking an independent LLM; or
2. inside Search, through AI-generated answers and zero-click experiences.

The people who still choose to visit the publisher are therefore plausibly a smaller but more motivated audience.

The data cannot prove that AI caused the full traffic decline, particularly because a domain migration occurred during the period.

However, Site 4 provides the strongest evidence in the portfolio consistent with the mechanism anticipated by the original forecast.

7. AI Visibility Does Not Equal AI Traffic

Across the observed sites, the ratio of AI-feature impressions to standard Web Search impressions ranged from approximately 8% to 32%.

The highest relative AI exposure did not belong to one of the GEO-optimized sites.

That does not imply that GEO was ineffective.

The ratio is strongly determined by query mix, market size and user intent.

Consumer-oriented ecommerce markets naturally generate larger quantities of informational, trust and comparison queries. Highly technical B2B websites operate within a much narrower prompt/search universe. Editorial publishers are naturally exposed to very large numbers of answer-oriented searches.

The relevant optimization question is therefore not:

“Which site has the highest percentage of AI impressions?”

but:

“Given the AI-search opportunity available to each website, is it increasing visibility and downstream business value within that opportunity?”

On that basis, the two optimized sites remain the strongest cases: both recorded substantial ranking improvements and improving Search outcomes following dedicated interventions.

8. What Happens After AI Visibility Matters More

The useful question is not simply whether optimization makes AI systems expose a brand.

It is:

Can LLM-oriented Search Optimization help preserve or increase valuable visibility, referral traffic and commercial outcomes as discovery becomes AI-mediated?

The observational pattern was:

Site	LLM/GEO optimization	Traditional Search outcome	Business outcome
Site 1	Yes	Clicks and rankings ↑	Organic traffic and revenue strongly ↑
Site 2	Yes	Clicks, impressions and rankings ↑	Traffic resilient
Site 3	No	Clicks ↓ despite rankings ↑	Organic revenue materially ↓
Site 4	No	Informational traffic strongly compressed	Organic sessions −53.5%

These websites are not experimental treatment and control groups.

They differ by geography, market size, intent, content footprint, commercial model and search demand.

Causality therefore cannot be claimed.

But the observation is strategically relevant:

The two properties receiving dedicated LLM/GEO optimization are also the two properties showing the clearest resilience and growth in traditional Search and organic business performance — not merely improvements in AI visibility metrics.

9. Direct LLM Referrals Are Only the Visible Tip of the Iceberg

Direct referral traffic from ChatGPT, Gemini, Perplexity and similar systems is now measurable.

But it is a poor proxy for total AI content consumption.

An LLM can:

• expose a brand;
• summarize content;
• influence consideration;
• resolve an informational need;

without generating any click.

Analytics sees none of these interactions.

This mismatch is particularly visible for the editorial publisher, which receives substantial exposure in generative search environments but relatively little direct AI referral traffic.

Therefore:

GA4 materially understates content consumption occurring inside AI environments.

AI referral sessions should be treated as the recoverable, measurable portion of an increasingly larger answer-engine ecosystem — not as a complete measurement of LLM impact.

10. What the Original Forecast Got Right

The original forecast correctly identified the structural direction of travel.

The discovery layer is fragmenting between:

• traditional Search;
• AI-enhanced Search;
• standalone LLMs.

The original analysis also explicitly anticipated that users would increasingly bypass conventional search engines altogether, rather than framing the issue only as Google AI Overview cannibalization.

That risk proved especially relevant for informational queries, where receiving an answer can substitute directly for visiting the original website.

The editorial case demonstrates that this concern was not theoretical.

11. Where the Original Forecast Did Not Match Reality

Given the information available in 2025, the principal limitation was the assumption of proportionality.

The practical model was approximately:

15% less AI-era Search → 15% less traffic → 15% less commercial value

The actual data show a more complex system:

Traditional Search Demand

minus

Demand displaced upstream to LLMs

×

Organic visibility / ranking share

×

CTR after zero-click and AI SERP effects

plus

Direct AI referrals

×

Conversion efficiency and value per visit

The four sites demonstrate how differently those variables can interact.

• Site 1 loses impressions while gaining clicks and revenue.
• Site 2 maintains near-flat CTR while ranking improvements expand total clicks.
• Site 3 improves ranking while losing search demand.
• Site 4 loses more than half its Organic Search volume while the quality of the remaining audience improves.

A single decline coefficient cannot adequately model all four.

12. The Most Important Finding: Intent Determines Exposure

The clearest pattern emerges when websites are classified by the reason the user needs the website.

Transactional Ecommerce

The user ultimately needs a provider to purchase, activate or manage a service.

An AI system can answer questions and influence provider selection, but usually cannot replace the final commercial action.

Observed result: high resilience.

B2B Lead Generation

AI can answer research and technical questions, but complex business requirements still require supplier evaluation, consultation, contact or implementation.

Observed result: Search remains resilient.

Secondary-Market Ecommerce

Transactional intent still offers protection, but smaller demand and a narrower query universe can create greater volatility.

Observed result: organic weakness despite improved SEO efficiency.

Editorial / Informational Publishing

The content itself frequently is the answer.

An AI system can therefore substitute directly for the page visit.

Observed result: by far the strongest traffic displacement.

This distinction is more actionable than a generic prediction that “LLMs will reduce SEO traffic by X%.”

13. A Revised Forecasting Framework

The original −15% / −25% / −35% curves should not necessarily be discarded.

They are still useful as portfolio-level stress scenarios.

But future models should treat AI disruption as a combination of separate variables:

Driver	Question
Traditional Search Demand	Is the underlying number of conventional searches expanding or contracting?
LLM Demand Displacement	How much demand may never reach Google or Bing?
Organic Visibility	Is the site capturing a larger share of the remaining search market?
AI-SERP Cannibalization	How much click propensity is lost after the search occurs?
Direct AI Referrals	How much traffic is recovered from LLMs and answer engines?
Traffic Quality	Are the remaining visitors more engaged or higher-intent?
Conversion Efficiency	Does each remaining visit generate more or less business value?
Paid Substitution	Can paid acquisition compensate for organic compression?
Business-Model Exposure	Can the AI answer substitute for the reason the user needs the website?

This framework explains the observed outcomes far better than a universal traffic-loss percentage.

Final Conclusion

AI disruption is real, but it is not a uniform traffic tax.

The original 2025 forecast correctly anticipated a fundamental shift in how people obtain information.

One year later, the evidence shows that the transition operates through multiple mechanisms:

some queries never reach Google because users start inside LLMs;
some searches occur but terminate inside AI-powered SERPs;
some AI interactions still generate website referrals;
and the economic value of the remaining visits varies dramatically by website type.

The resulting evidence is more sophisticated than either “the forecast was right” or “the forecast was wrong.”

For the GEO-optimized transactional ecommerce and B2B lead-generation properties, the expected absolute decline did not occur. Search visibility and traffic proved resilient, while measurable commercial performance improved substantially in the ecommerce case.

For the non-optimized secondary-market ecommerce property, organic deterioration occurred despite improving SEO positions, showing that traditional search-demand contraction can happen independently of ranking performance.

For the informational publisher, the disruption was substantially larger than the original −15% downside scenario, while the remaining organic audience became more engaged.

The revised strategic thesis is therefore:

The impact of LLMs is determined less by how much AI is present in Search and more by whether AI can substitute for the user’s reason to visit the website.

For transactional and relationship-driven websites, AI can alter discovery without eliminating the underlying customer journey.

For informational publishers, the answer itself may be the product — and therefore the click is structurally far more exposed.

That is the distinction the original 2025 model could not yet make.