Findable but forgettable: are Hungarian companies ready for AI search?

A machine-readiness audit of 15 Hungarian firms · Budapest · June 2026

Abstract

As search shifts from a ranked list of links to a synthesised AI answer, the operative question for a business changes from can we be found? to are we known? On a hand-built sample of fifteen Hungarian companies audited in June 2026, most firms are findable but not "known": when an AI searches the live web, they surface — but with the web switched off, the model does not reliably recall them, and over half have no authoritative entity record for a model to anchor to. The healthy part of the funnel is retrieval; the gaps are recall and authority — precisely the positions that AI answers make more durable and harder to buy. A second, blunter obstacle sits underneath all of it: two-thirds of the sites refused a basic automated crawler outright.

This is a small, single-country, indicative study, not a national statistic. Its value is the pattern, not the percentage.

1. From "ten blue links" to the answer itself

Classic search engine optimisation was built for one job: rank a page high enough in a list that a human clicks through. AI Overviews and chat assistants change the shape of the result. Instead of a list to choose from, the user gets a single composed answer that cites a small number of sources — and often resolves the question without a click at all.

That reshapes what visibility means. It helps to separate three tiers:

• Being seen (retrieval). When an AI searches the live web for a category, does the company surface among the sources?
• Being known (recall). With the web off, does the model already carry the company in its trained memory?
• Being trusted (authority). Is there an authoritative record — a Wikipedia/Wikidata-grade entry or knowledge panel — that the model can anchor to and corroborate?

Classic SEO optimised the first tier. AI answers collapse "being seen" into the answer itself, which raises the value of the second and third tiers: recall and authority are slower to build, harder to fake, and therefore more defensible. The key distinction this study keeps returning to is that being machine-readable is not the same as being machine-known. A site can be immaculately structured and still absent from the model's memory; a brand can be famous and recalled, yet have nothing structured for a model to anchor to.

2. Method, in brief

Fifteen Hungarian firms were audited on 2026-06-30, spread across three size tiers (five large, five medium, five small) and a range of sectors. Each firm was scored on two axes, every signal on a simple 0/1/2 scale:

• Readability — how easily a machine can parse the site: server-rendered text, structured data, crawlability, canonical/hreflang, clean headings, contact and product data. Collected only for the sites that actually loaded for the crawler.
• Credibility — how much reason a machine has to trust and recall the firm: an authoritative record, name disambiguation, third-party corroboration, cross-source consistency, named people, plus two live probes.

The two probes matter most:

• Open-book (retrieval): with web search on, does the firm surface for a realistic category query (scored prominent / partial / absent)?
• Closed-book (recall): with web search off, does the model recall the firm from memory (accurate / partial / no recall)?

Read the limitations in §8 before quoting any figure. In particular: the probes are single-shot and from one model, "retrieval" used an AI/web-search surface rather than a live Google AI Overview, and readability exists for only five of the fifteen sites.

3. What the audit found

3.1 Retrieval is the healthy part

Findability is not the problem. Eleven of the fifteen firms surfaced prominently in a live AI/web-search for their category, and only two were outright absent (the remaining two appeared partially). Whatever is broken here, it is not that the web has never heard of these companies.

3.2 Recall is the gap

Switch the web off and the picture inverts. In a single closed-book probe, the model recalled only three of fifteen firms accurately; eight were partial, and four drew a blank. Most strikingly, four of the five largest national champions landed on "partial," not "accurate." Size and fame buy some memory, but far less reliably than one might assume — and recall, not retrieval, is where the sample is thin.

(Recall is also the noisiest signal in the study; see §8. Treat these as indicative directions, not verdicts on any single firm.)

3.3 Authority is the size-graded gap

The clearest, most structural finding: seven of fifteen firms have no authoritative entity record at all — no Wikipedia, Wikidata, or knowledge-panel presence for a model to anchor to. And unlike recall, this gap is cleanly graded by size. On a 0–2 scale, average authority runs 2.0 for the large firms, 0.8 for the medium, and 0.4 for the small. The smaller the firm, the less likely a model has any trustworthy structured anchor for it.

Averages on a 0–2 scale. Retrieval and recall barely separate medium from small; authority does.

3.4 The locked front door

Before any of the above, there is a cruder barrier. Ten of the fifteen sites returned an HTTP 403 to a plain automated fetch — a Cloudflare/WAF-style bot block. A site that refuses a generic crawler is, by the same mechanism, harder for AI ingesters to read. And the block was total among the smallest firms: five of five. The single most common machine-readiness failure in the sample is not bad schema; it is a door that won't open.

3.5 Readability is a choice, not a function of size

Among the five sites that did load, machine-readability ranged wildly — and did not track company size. At one end, the largest bank's site scored a perfect 14/14 (clean server-rendered content, valid structured data). At the other, a mid-size software firm scored 4/14. Same country, same crawl, opposite outcomes. Good machine-readability is built, not inherited; it is a per-site decision, not a by-product of scale. (Because only five sites loaded, there is no honest whole-sample readability average — see §8.)

3.6 The SME pattern: corroboration without a record

It would be easy to assume small firms are invisible because nobody talks about them. The data says the opposite. Independent third-party corroboration and cross-source consistency were near-maxed across every size tier — even a roughly nine-person creative agency and a ~27-room hotel were talked about consistently. What the smaller firms lack is not buzz; it is a structured, authoritative record to convert that buzz into something a model can anchor and trust.

4. The two-by-two: authority and recall are different muscles

The richest way to read the sample is to cross authority (is there a record?) against recall (does the model remember?). Two anonymised cases sit in opposite corners:

• Fame without a footprint. An internationally recognised fashion house surfaces prominently and was, in a single probe, recalled accurately from memory — yet has no brand-level entity record (only its founder has a wiki page). Strong on recall, exposed on authority: there is nothing structured for a model to anchor to if its memory drifts.
• A footprint without fame. A mid-size software firm has a substantive Hungarian Wikipedia article (authority present), yet was absent in a live "custom software in Budapest"–style search and drew a blank from memory — and posted the lowest readability of the loadable sites (4/14). A real record can still go unseen if retrieval and on-page readability don't carry it.

The lesson: authority and recall do not come as a bundle. You can have either without the other, and a complete position needs both — plus the readability and crawlability to make them legible.

5. Four short cases

"Readability is a choice." The largest bank's site scored 14/14; a mid-size software firm's scored 4/14. Identical crawl, opposite results — machine-readability is engineered, not granted by size.

"The record nobody surfaces." The same software firm shows how a genuine Wikipedia-grade record can still fail to convert: with weak on-page readability and no retrieval pickup, the record exists but does no work.

"Fame without a record." The fashion house is the inverse failure mode — remembered and retrieved, but with no structured brand entity underneath the fame.

"The local-language blind spot." A leading premium winery has a dedicated Hungarian-language Wikipedia entry — but an automated check looking only at English-language and Wikidata signals initially scored its authority at zero. The record was real; the detector was English-centric. This is both a finding about the firms and a caution about the method (see §8): authority detection is brittle across languages, and a CEE audit that reads only English will systematically under-credit local- language presence.

6. What this implies for companies

For the smaller and mid-size firms, the instinct to "get more press" is misdirected — the sample already has corroboration. The missing layer is structured authority and an open door:

1. Unlock the front door. A site that 403s a generic crawler is invisible to the same machinery that feeds AI answers. This is the cheapest, highest-leverage fix, and it is most acute for the smallest firms.
2. Create a structured, authoritative record. A substantiated Wikipedia entry where notability supports it, a Wikidata item, and sameAs links wiring the official profiles together give a model something to anchor and corroborate — the difference between being talked about and being known.
3. Make the on-page layer machine-readable. Server-rendered content and valid structured data are a per-site decision; the audit shows firms of every size getting this right or wrong independently of scale.

For the large firms, retrieval and authority are largely handled; the soft spot is recall. Even national champions were only partially recalled from memory in a single probe — a reminder that training-time presence is earned over time and across sources, not bought at launch.

7. Why recall and authority are the durable positions

The mechanism behind the framing is worth stating plainly. Classic SEO competed for rank in a list, where many results could share the page and a click was the prize. AI answers compress that: a single synthesised response cites few sources, frequently resolves the query without a click, and leans on the model's understanding of entities rather than the keyword match of a page. Ranking, citation, and recall become three different things.

The supporting literature through 2024–2026 points the same way — work on zero-click search (Pew), analyses of AI-citation patterns and "ghost citations" (Ahrefs, Semrush), 2026 field and traffic data on referral erosion (an SSRN field experiment alongside SparkToro and Chartbeat), and academic work on generative-engine optimisation (Princeton, KDD 2024). These are dated and, where they come from vendors, should be read as attributed evidence rather than neutral fact. But the direction is consistent: as answers replace lists, being cited and recalled as a trusted entity outlasts ranking a page. That is why this study weights recall and authority — they are the slower-built, harder-to-buy positions that AI search rewards.

8. Limitations — read before quoting

This study is deliberately modest about what it can claim. Its credibility depends on these caveats:

• Small, single-country, convenience sample (n = 15, Hungary). Indicative, not representative. Every figure here describes this 15-company sample, never Hungarian companies at large.
• Probes are single-shot (k = 1) and from one AI model. Recall in particular is noisy — on different single runs the same winery scored both "accurate" and "no recall." Treat recall as indicative ("in a single probe," "tended to"), not as a hard fact about any firm.
• Readability exists for only 5 of 15 sites; the other ten blocked the crawler. There is no honest whole-sample readability average — only those five are measured.
• "Retrieval" used an AI/web-search surface, not a live Google AI Overview. AI Overviews were not default-on in Hungary at audit time. Read these as "AI/web search," not "Google's AI Overview said…".
• Authority detection is imperfect, especially for local-language records (see the winery in §5). The authority figures use a grounded manual re-check; "authority" means "a Wikipedia/Wikidata-grade record," not an absolute.
• No fabricated quotes. Nothing here attributes a manufactured statement to an AI about any named firm; only the probe outcomes are real.

9. Appendix — the sample (anonymised)

Companies are described, not named. Two columns are shown because they are internally exact across the sample: whether the site loaded for a basic crawler, and whether an authoritative record exists. Retrieval and recall are reported only as the aggregate distributions below, which are the study's exact figures.

Aggregate distributions (n = 15).

• Retrieval: prominent 11 · partial 2 · absent 2.
• Recall (single probe): accurate 3 · partial 8 · no recall 4.
• Authority: record 8 · none 7.
• Sites that loaded for the crawler: 5 · blocked (403) 10.
• Authority by size (0–2): large 2.0 · medium 0.8 · small 0.4.

Audited 2026-06-30. Indicative findings from a 15-company Hungarian sample; see §8.