AI search in 2026: the drop,
the outage question, and
how to get cited.

As of 31 July 2026, there is no confirmed widespread Google AI Overviews outage. That is the direct answer, and it is worth stating plainly because the search volume behind questions like "is AI overviews down" spikes far more often than the feature actually breaks.

What has genuinely happened in this window is narrower and more technical. Google's developer-facing AI infrastructure — AI Studio and the Gemini API — has recorded intermittent incidents through June and July 2026, including an infrastructure outage that Google logged and investigated, plus problems with streaming Deep Research responses. Third-party status aggregators such as StatusGator and Entireweb have been carrying user-submitted outage reports for Gemini and AI Studio at a low, steady rate throughout the period.

None of that is the same system as AI Overviews. When Gemini went down in June 2026, Search Engine Roundtable reported that Google Search's generative surfaces — AI Overviews and AI Mode — were unaffected. That is the single most useful fact in this section: the assistant product and the search feature are decoupled. A Gemini incident trending on X is not evidence that AI Overviews is broken, and treating the two as one system is the most common way people convince themselves there is an outage when there is not.

Why "AI Overviews is down" feels true even when it isn't

Four things produce the sensation of an outage without an outage existing.

Query-level suppression. Google does not attempt an AI Overview on every search. It withholds them on queries it classifies as sensitive — medical specifics, legal advice, elections, some financial topics — and on queries where its confidence is low. Google tunes these thresholds continuously. When a threshold moves, an entire category of queries can lose its AI Overview overnight. To anyone watching that category, it looks exactly like a failure.

Personalisation and locale. AI Overviews vary by country, language, account state, device and experiment bucket. Two people running the identical query in the same city can legitimately see different results. Before concluding anything is broken, run the query logged out, in an incognito window, and ideally from a second network.

Ordinary ranking movement. If your page was cited inside an AI Overview last week and is not this week, the overview did not break — you lost the citation. This is by far the most common real cause, and it is the one people are least willing to consider, because "Google is broken" is a more comfortable explanation than "a competitor wrote a better answer."

Tracking blind spots. Conventional rank trackers report a fixed, largely US-centric, daily-sampled, blue-link keyword set. They were not built to watch generative surfaces. Analysis of 2026 volatility has repeatedly noted that as AI surfaces absorb more of the search experience, the share of consequential movement that blue-link trackers cannot see keeps growing. A tracker showing a flat line during a week when your AI citations collapsed is not reassurance; it is a measurement gap.

The last item is the one most teams cannot do, and it is why "was there an outage?" is so often unanswerable after the fact. Generative answers are not archived anywhere by default. If you did not record what the engine said yesterday, you have no baseline today, and every change looks like either nothing or a catastrophe. The whole discipline of generative engine optimization depends on keeping that evidence.

What is genuinely unstable in 2026

Volatility and outage are different claims, and the volatility claim holds up better. Through 2026 Google has continued shipping core and spam updates on its usual cadence, and independent trackers logged elevated ranking movement through March, April and June 2026 before conditions calmed. Separately, AI Overview presence itself has been measured as remarkably stable — hovering around 32.5% of monitored queries and moving only between roughly 31.0% and 34.4% across measurement periods. Read together, those two facts say something specific: the surface is stable, the contents of it are not. AI Overviews are not going away, but who gets cited inside them changes constantly.

That is the honest state of play. If you came here because AI Overviews disappeared from a query that matters to you, the answer is almost certainly not an outage — and the fix is not to wait for Google, it is to find out whether your page is still retrievable, still quotable, and still the best available answer. The rest of this page is about how to do that.

"AI took the clicks" is true but not actionable. Underneath it are eight distinct mechanisms, and they call for different responses. Knowing which ones apply to you is most of the work.

1. Answer sufficiency

The dominant mechanism. The generated summary fully resolves the query, so there is no reason to click. This hits definitional and procedural content hardest: "what is DNS", "how many ounces in a cup", "symptoms of X". If a competent two-paragraph answer ends the user's need, the page that used to earn that visit is now redundant. Response: stop competing on facts that fit in a paragraph. Move up the value chain to content that requires your judgement, your data, or your product.

2. Vertical displacement

Even when a user does want to click, the AI Overview physically occupies the space where the first organic result used to be. On mobile especially, position one can now sit below the fold. This is why measured CTR declines are steepest at position one — the rank did not change, the geography did. Response: being cited inside the overview is worth more than being first below it, which inverts the usual optimisation target.

3. Citation compression

A results page offers ten organic slots. A generated answer typically names three to five sources. That is a structural reduction in how many sites can win a query, independent of quality. The middle of the distribution — pages that were good enough for position six — loses the most, because there is no position six inside an answer. Response: a strategy of "rank respectably for many terms" degrades badly here. Depth on fewer questions beats breadth.

4. Migration of informational intent to assistants

Some queries never reach Google at all now. They are typed into ChatGPT, Perplexity, Claude or Gemini directly. Google's own engagement data cannot show you this, because the search never happened there. This is the invisible portion of the decline, and it is why measuring only Google understates the problem. Response: track prompts across engines, not keywords in one engine.

5. Trust erosion after visible failures

High-profile AI answer errors have made a meaningful segment of users sceptical of generated summaries. Some scroll past them deliberately. This partially offsets mechanism one and is one reason CTR estimates vary so much between studies — user populations differ in how much they trust the box. Response: this is a genuine opportunity. Being the cited source under a summary users distrust is worth more than it looks, because the sceptical clicker is a high-intent clicker.

6. Query fragmentation

Conversational interfaces encourage longer, more specific, multi-turn questions. One old query becomes three new ones, each with lower individual volume. Keyword tools built on aggregate volume systematically under-report this tail, so demand appears to shrink when it has actually dispersed. Response: a keyword gap analysis needs a prompt-gap companion, or you will keep optimising for the shrinking head.

7. Brand-name shortcutting

As assistants recommend a shortlist, more users skip discovery entirely and search the brand they were told about. This shows up as declining engagement on category and comparison terms and rising branded search. It is not lost demand; it is demand that moved upstream into the assistant's recommendation. Response: if you are not on the shortlist, you never enter the funnel. This is exactly what ChatGPT ranking work is for.

8. Measurement blind spots

Some of the reported decline is an artefact. Referrals from AI surfaces are attributed inconsistently across analytics setups, blue-link rank trackers cannot see generative movement at all, and Search Console reports impressions for AI surfaces in ways that are easy to misread. Teams routinely report a collapse that is partly a reporting change. Response: before diagnosing a decline, confirm your measurement can see the thing you are measuring. A technical audit plus a clean look at AI referral sources will separate the real losses from the accounting ones.

An AI search optimization platform — commonly called a GEO platform — runs a set of buyer questions against AI assistants on a schedule, records whether your brand is named and which sources are cited, compares that against competitors, and tells you what to change so you get quoted more often.

The category appeared in under two years and is already crowded, partly because the first job is easy to build and the second is not. Anything can query an API and count brand mentions. Very little can tell you why the mention did not happen and hand you the fix.

The five jobs

1. Prompt discovery. Keywords are not prompts. "crm software" is a keyword; "what's the best CRM for a 12-person nonprofit that needs Xero integration" is a prompt. A platform should generate a realistic prompt set from your site and category, then let you edit it, because the prompt set determines everything downstream. A bad prompt set produces a confident, worthless score.

2. Multi-engine scanning with retained evidence. The scan must hit each engine for real and store the full answer text. This is the single most important technical distinction in the category. Generated answers vary between identical runs, so a score without stored evidence cannot be audited, cannot be trended honestly, and cannot settle an argument with a stakeholder. If a vendor cannot show you the raw response behind a number, the number is a claim, not a measurement.

3. Competitive and citation analysis. Being absent matters less than who is present instead. The useful output is: which competitors are named, in what order, and which specific URLs the engine cited — including the third-party pages, review sites and forums doing the work. Those cited domains are your target list.

4. A technical retrievability audit. This is where most tools stop and where most of the actual problem lives. The reasons an engine skips a brand are usually mechanical: the answer only exists after JavaScript runs, an AI crawler is blocked at the CDN, the page answers six questions and therefore none of them cleanly, the schema contradicts the visible text. A dashboard reporting 0% share of voice while a firewall rule blocks GPTBot is an expensive way to learn nothing.

5. Content production that closes the gap. Every missing answer is a page you have not written. The platforms worth paying for turn each gap into a draft structured to be extracted — direct answer first, question-led headings, quotable statistics, matching schema.

The distinction that matters when you are choosing

Ask a single question of any vendor: when I am not mentioned, does this tell me why, and does it give me the page that fixes it?

If the answer is no, you are buying a monitor. Monitors have real value — you cannot manage what you cannot see, and knowing a competitor owns your category inside ChatGPT is worth knowing. But a monitor produces a chart and a feeling of urgency, then leaves the work to you. A platform closes the loop.

There is a reason people search for "SEO platforms" when the goal is an AI citation rather than "GEO platforms", and the instinct is correct. You cannot be cited from a page that was never retrieved, and retrieval runs on the same signals organic ranking always has. A pure AI-monitoring tool with no SEO underneath it is measuring the last ten yards of a race you are not in.

So the real question is not which category to buy from. It is which tool does both jobs without dropping either.

The two-jobs test

A platform that gets you cited in ChatGPT has to handle a chain that runs: be crawlable → be retrievable → be relevant → be quotable → be corroborated → be measured. Classic SEO suites own the first three links well and the last three barely. Pure GEO monitors own the last one and none of the others. The gap in the middle — quotability and corroboration — is where citations are actually won and lost, and it is the least-served part of the market.

How to choose in practice

If you have no AI measurement at all, start with a free scan before you buy anything. Most teams discover something in the first run that reorders their priorities — a competitor named as the category default, or an engine confidently stating pricing that has been wrong for a year.

If your SEO fundamentals are weak, fix those first, and pick a platform that reports both. Buying AI visibility tracking for a site that renders its content client-side is paying to watch a problem you could have fixed. Start with the on-page checklist and the technical audit guide.

If you already have a mature SEO suite, the add-on from that vendor is the path of least resistance and a defensible choice — right up to the point where you need to know why you are not being cited. That is when suites run out.

If you are an agency, engine breadth and client reporting matter more than depth on any one site, which pushes you toward Authoritas or Profound.

If you want one tool that covers the whole chain, that is the case for Klepha, and it is the reason it sits at the top of the ranking above.

Every quantitative claim above comes from a named, dated, third-party source rather than from Klepha's own product data. Where sources disagree — and on click-through impact they disagree considerably — we have given the range and explained the methodological reasons rather than picking the most dramatic figure.

Two deliberate omissions are worth flagging. We have not invented a July 2026 AI Overviews outage, because the public record does not support one. And we have not published precise competitor pricing beyond what vendors state publicly, because it changes frequently and a stale number is worse than a band.

This page is maintained. It carries a visible last-updated date and will be revised as the underlying research moves.

Sources

• Pew Research Center — study of approximately 68,000 real search queries; clicks fell from 15% to 8% of visits when an AI summary was present.
• Ahrefs — CTR impact of AI Overviews; 34.5% decline for position one (April 2025), approximately 58% (February 2026).
• Search Engine Roundtable — reported that the June 2026 Gemini outage left Google Search's AI Overviews and AI Mode unaffected.
• Google — 2.5 billion monthly active users for AI Overviews and 1 billion for AI Mode, stated on the Google blog, 19 May 2026.
• Moz (Tom Capper) and Similarweb — early AI Mode adoption at roughly 1–2% of US search traffic; Nick Fox, VP of Search, cited 75 million AI Mode users in December 2025.
• Vercel and MERJ — analysis of 500M+ GPTBot fetches finding zero JavaScript execution; GPTBot downloaded JS files ~11.5% of the time without running them; same behaviour observed for ClaudeBot and PerplexityBot.
• Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande — "GEO: Generative Engine Optimization", ACM SIGKDD 2024 (arXiv:2311.09735); Statistics Addition, Quotation Addition and Cite Sources as top-performing methods, ~30–40% relative visibility improvement across ~10,000 queries.
• SERP feature tracking, 2026 — AI Overview presence averaging ~32.5% of monitored queries within a 31.02%–34.40% band; ranking volatility down roughly 65% in one measurement period from late March 2026.
• Vendor documentation and published pricing, July 2026 — Profound, Semrush, Ahrefs, Peec AI, Otterly.ai, Authoritas, SE Ranking.