If your website’s organic traffic fell between 2024 and 2026 and you cannot find the competitor who took it, there may not be one. A structural change in how search results are presented has removed a large share of clicks from the web entirely. The visits did not move. They stopped happening.

This has produced two bad responses among Nigerian businesses. The first is to ignore it, on the reasonable-sounding basis that Nigerian customers still search normally. The second is to buy an expensive package of AI optimisation services from a vendor whose methods nobody can verify. Both are mistakes, and the evidence published over the last eighteen months is now good enough to say why.

How large is the change, really?

The most methodologically careful measurement comes from the Pew Research Center, which tracked the actual browsing behaviour of a panel across roughly 68,000 real Google searches. It found that users clicked an external link on 8 per cent of searches where an AI summary appeared, against 15 per cent where none did. Sessions ended entirely on 26 per cent of pages carrying an AI summary, against 16 per cent otherwise.

Other measurements point the same way with different magnitudes:

• A randomised field experiment reported in April 2026 found AI Overviews reduced organic clicks on triggered queries by about 38 per cent, while self-reported satisfaction with search was essentially unchanged. Removing top-of-page AI Overviews roughly doubled outbound clicks.
• Ahrefs, analysing Search Console data, reported click-through rate falling by around 58 per cent for top-ranking pages when an AI Overview appeared. Seer Interactive measured a comparable 61 per cent drop across billions of impressions.
• BrightEdge put AI Overview trigger rates at roughly 48 per cent of tracked queries by early 2026, up sharply year on year. SparkToro’s clickstream analysis for January to April 2026 put the overall zero-click rate near 68 per cent.
• Google announced at its 2026 developer conference that AI Mode, its fully generative interface, had passed a billion monthly users. Independent clickstream data over the same window put AI Mode at only about 0.34 per cent of searches, which suggests the larger disruption has not arrived yet.

These studies use different panels, query sets and methods, and their estimates range from mild to severe. Treat any single figure with caution. The direction, though, is consistent across every independent measurement published, which is more than can be said for most claims in this field.

One finding deserves separate emphasis because it changes the strategy. Seer Interactive found that brands cited inside an AI Overview received roughly 35 per cent more organic clicks and 91 per cent more paid clicks than uncited brands appearing on the same results page. The market is not shrinking uniformly. It is splitting into cited and bypassed, and the gap between those two states is now larger than the gap between position one and position five used to be.

Does this land the same way in Nigeria?

Partly. Nigeria had roughly 109 million internet users at the end of 2025, an online penetration of about 45.5 per cent, against 165 million active cellular connections. The NCC put active internet subscriptions at 142.6 million as of October 2025. This is an overwhelmingly mobile market, and roughly four in five AI Overview queries globally occur on mobile, so Nigerian users are squarely inside the affected population.

But the exposure is uneven, and the distinction matters for where you spend.

Local transactional queries are relatively protected. Someone searching for a solar installer in Lekki, a diagnostic centre in Wuse or a school fees portal needs an address, a phone number or a login. The answer is a destination, not a summary. These queries still produce clicks, and for most Nigerian SMEs they are the commercially important ones.

Informational and comparison content is heavily exposed. Guides, definitions, how-to articles and product comparisons are exactly what generative answers are best at absorbing. If your traffic strategy was built on ranking for questions, that strategy has been substantially devalued.

There is a Nigerian opening that few are taking. Generative systems answering questions about Nigeria are drawing on a thin, often outdated and frequently foreign-authored source base. Ask an assistant about Nigerian regulatory deadlines, local pricing or sector-specific practice and the answers are noticeably weaker than the equivalent for the United States or United Kingdom. That gap is an opportunity for any Nigerian organisation willing to publish primary, dated, verifiable local information. It is easier to become a cited source on a Nigerian topic today than it will be in three years.

What does the evidence say does not work?

This is the part of the discussion where most money is currently wasted. On 15 May 2026 Google published its first official guidance on optimising for generative features in Search. Its position is blunt: these features run on the same core index and ranking systems as ordinary Search, so optimising for them is still SEO. The guidance then names specific tactics site owners can stop worrying about.

• txt and similar machine-readable files. Google states plainly that it does not use them. Google may crawl such a file, but it receives no special treatment.
• Chunking content into small blocks for machine consumption. Google says its systems handle multiple topics on a page and surface the relevant part.
• Rewriting content into an AI-friendly voice. The systems understand synonyms and meaning.
• Special schema markup for AI features. Not required.

Independent testing points the same way. An Ahrefs study published in May 2026 measured AI citations across 1,885 pages against a control group of about 4,000, before and after adding schema markup, and found no citation lift — with a small decline on Google’s AI Overviews. The caveat matters: the tested pages already had strong citation baselines, so the result speaks most directly to brands with existing AI visibility. Separately, an analysis of roughly 500 million AI bot visits over ninety days recorded only a few hundred actual fetches of llms.txt files, with adoption across domains flat at around 10 per cent.

The honest reading is narrower than either camp claims. Schema markup remains worth implementing, because it drives rich results in conventional search and costs little. It is simply not an AI visibility lever, and should not be sold as one. An llms.txt file costs nothing to publish and may serve systems other than Google, so publish it if you like. Do not pay a consultant for it.

What does the evidence say does work?

The strongest evidence in this field is a peer-reviewed paper presented at ACM KDD in 2024 by researchers from Princeton, Georgia Tech, the Allen Institute and IIT Delhi, which tested content interventions across thousands of queries and several generative engines. It found that restructuring content to include cited sources, direct quotations from authorities and concrete statistics produced visibility gains in generated answers in the range of 30 to 115 per cent. Two interventions that the marketing industry sells heavily did nothing: rewriting purely for fluency produced almost no lift, and keyword stuffing produced a small negative effect on some engines.

Two qualifications from the same research are usually omitted when it is cited. The same intervention can help substantially on one engine and barely register on another, because retrieval and ranking models differ. And the gains, while real, are bounded — no strategy doubled citation rates across the board. This is a discipline of small compounding wins, not a lever.

That leaves a short list of things worth doing, all of which happen to also be good writing:

1. Answer the question in the first sentence of each section, then give the context. A passage that only makes sense after three paragraphs of preamble cannot be extracted.
2. Make sections self-contained. If a paragraph relies on a pronoun referring to something two screens above it, it will be retrieved without its meaning.
3. Attribute every statistic to a named source with a date. This is the single most consistent finding across the research, and it is also the thing most Nigerian business blogs do least.
4. Shape headings as the questions your customers actually ask, in their words, not your internal product vocabulary.
5. Keep facts current and date-stamp updates. Generative systems favour recency, and stale figures are the fastest route to being dropped from a cited source pool.
6. Keep entity information identical everywhere: business name, address, phone number, description. Across your site, your Google Business Profile, your directory listings and your social profiles. Inconsistency here is why assistants describe businesses inaccurately.
7. Publish something nobody else has. Original data, your own operational numbers, a survey of your customers, a price index for your sector. Google’s own guidance now leans explicitly on content that is not a commodity restatement of what is already online.

The technical floor

None of the above matters if the systems cannot read your pages. Check these first, because they are cheap, verifiable and frequently broken on Nigerian websites:

• Your robots.txt does not block the crawlers you want citing you, and your CDN or web application firewall is not silently rejecting their requests.
• Important content is rendered server-side, not assembled by JavaScript after the page loads.
• Nothing commercially important sits behind a login, a pop-up or an interaction that a crawler cannot complete.
• Heading hierarchy is real: one H1, meaningful H2s, no headings used purely for visual size.
• Your pages load fast enough to be crawled efficiently, which on a Nigerian-hosted or Nigerian-audience site usually means fixing hosting and image weight before anything else.
• Google Search Console is connected, so you can use its generative AI performance report rather than guessing.

How do you measure this without buying a dashboard?

Measurement in this field is genuinely poor, and a large part of the vendor market exists to sell certainty that does not exist. A workable free approach:

• Use the generative AI performance report in Search Console for Google surfaces.
• Segment referral traffic from assistant domains in your analytics. Volumes will be small; the conversion quality is usually high, because someone who clicks out of an AI answer is deep in an evaluation.
• Watch branded search volume and direct traffic. If you are being described accurately in AI answers without a link, these are where the effect surfaces.
• Keep a fixed panel of twenty prompts a real customer might ask, and run them monthly across the major assistants. Log whether you appear, whether the description is accurate, and who appears instead. This is crude, manual and more informative than most paid tools.

The case for doing much less than the industry recommends

An honest article should make the opposing argument properly, so here it is.

For most Nigerian SMEs, absolute AI referral volume is still small. Conventional Google organic search, WhatsApp, Instagram and word of mouth remain far larger sources of business, and a naira spent improving conversion on the traffic you already have will usually beat a naira spent chasing citations. The generative optimisation services market is growing at rates that should make any buyer suspicious, and much of what it sells is unfalsifiable by design. Google’s own position is that this is ordinary SEO, which implies a business with slow hosting, thin content and no Google Business Profile should fix those first and ignore the acronyms entirely.

There is also a reasonable argument that the trend is being over-extrapolated. Independent clickstream data showed the United States zero-click rate falling slightly between December 2025 and March 2026 on a strict methodology, and AI Mode remains a rounding error in query share. People predicting the end of the web have been wrong before.

Weighing it up: the defensible position is that AI visibility should be a by-product of publishing genuinely useful, original, well-sourced, crawlable content, not a separate budget line with its own vendor. Everything on the evidence-backed list above improves your conventional search performance too. That is the test to apply to any proposal you receive. If a tactic only helps if the AI thesis is correct, it is a bet. If it helps either way, it is a decision.

The bottom line

The measured decline in clicks is real, substantial and unlikely to reverse. The market is separating into sources that get cited and sources that get bypassed, and the citation premium is large. But almost none of the tactics being sold to close that gap survive contact with evidence, and the search engine at the centre of it has now said so in writing.

What survives is unglamorous: be crawlable, be structured, be specific, be current, be attributable, and publish something about Nigeria that nobody else has published. That was good advice in 2019. It is now the only advice with data behind it.

Statistics in this article were current as of July 2026. This field changes monthly; verify before relying on any figure.

Sources

Google Search Central, optimising for generative AI features in Google Search;  Search Engine Journal, Google’s AI search guide calls AEO and GEO still SEO; Search Engine Journal, randomised field study on AI Overviews and organic clicks; Search Engine Land, Google zero-click searches in early 2026; Similarweb, zero-click marketing and the 2026 data;

Refinea, operational review of generative engine optimisation evidence including the Ahrefs schema study; We The Flywheel, review of the ACM KDD 2024 generative engine optimisation paper; DataReportal, Digital 2026: Nigeria; AllAfrica, NCC data on active internet subscriptions in Nigeria