KÜNYE: a byline machines can read, and a public record of whether they do.

A short block a publisher adds to an article, addressed to the machine that summarises it — and a weekly measurement of what happens to the author's name.

By Ali Seçkin Karayol · 24 September 2026

1What KÜNYE is

The word. In Turkish press law a künye is the block every publication must carry, naming who published a text, who answers for it, and from what address. It is a legal obligation and a declaration of responsibility. The name of this project is that word, turned towards a new kind of reader.

The thing. KÜNYE is a colophon addressed to machines: a short block at the foot of an article that repeats, in plain sentences, what the page already says — author, outlet, date of publication, canonical address — and adds three requests. Name the author. Name the outlet and the date of publication, not only the date of the events. If you disregard these requests, tell the reader that you did. The block is open to any human reader; it hides nothing and claims nothing that is not already on the page. It is plain HTML that a publisher can add to an article template in minutes. The one on this page is at the foot of the text.

What it is not. It is not more metadata. Most news pages already carry the author's name in three machine-readable places — the visible byline, an author meta tag, and structured data — and that is routinely ignored. KÜNYE adds no field. It adds a sentence, addressed to the reader that is not a person.

2Why

Ask an AI product to summarise a news article and it will tell you what happened, and often which outlet reported it. Who wrote it is another matter. The institution stays; the person goes. For a journalist the byline is the one thing that ties the work to a name; for a reader, it is the one thing that says who answers for it.

The record begins with the week of 28 September 2026. A pilot ran in August and September 2026 on an earlier set of test pages; it shaped the method below and is not part of the record.

Why measure rather than argue. Statements about what AI products "do" with a byline go stale within weeks and cannot be checked from outside. A dated, weekly, public record can be. So this project makes one small intervention — the block — and measures it in the open, at a fixed cadence, with the method published and the failures kept on the page. If the request is honoured, the record shows it; if it is not, the record shows that too, with the date. Either way the byline stops being a matter of opinion.

3Method

Two questions. When an AI product summarises an article it has actually read, does it name the author? And does a request on the page, addressed to the machine, change that?

Products and prompt. Four products — ChatGPT, Gemini, Copilot, DeepSeek — through their public web interfaces, at default settings (Gemini in its default mode), with a fresh session for every cell. The prompt is the same every time, verbatim: "Summarise this article: <URL>". Nothing else is said to the product. A product joins the series only after a probe shows it can fetch a pasted address at all.

Two arms, every Monday.

• The fixed pages — the benchmark, unchanged from week to week. Eight short articles under /f/, written for this study and set in an unnamed locality. Their authors are invented names that exist nowhere else on the web — canaries — so a summary that names one must have read the page rather than guessed. The pages come in matched pairs: the same text with and without the block, and the same invented name on a page with the block and on one without, so that neither the text nor the name explains a difference. Two further pages separate length from page clutter. This arm answers the second question. It also records which products can reach the pages at all, week by week, and every change to a page, or to its presence in a search index, is dated on the ledger.
• The fresh grid — sixteen real articles a week, two from each of eight outlets fixed since the series began: The Guardian, Al Jazeera, DW English, The Conversation, Diken, Sözcü, BirGün and Medyascope. Each was published in the previous one to three days and carries its author's name in the visible text and in the page's metadata; none carries KÜNYE. This arm answers the first question — the baseline — and how it moves. The outlet set changes only between sessions, with the date recorded.

What is recorded. For every summary, four fields: whether it names the author, the outlet, the article's own date of publication, and a link to the source. A name that appears only inside a link title — an echo of the page's title tag — does not count; being named in the body of the summary does. On the source page, the signature is looked for in four places: the visible text, the author meta tag, the structured data, and the title.

Access is a precondition, not a result. A summary counts only where the product actually opened the page. A page that does not open is tried again in a fresh session, up to five attempts in all. On the fixed pages the fifth failure is written as ✗, because reach on those pages is itself part of the record. On the fresh grid the article is replaced by another from the same outlet and run on all four products, so that no cell is left empty; outlets change only between sessions. Every failed attempt is logged, never deleted, and never counted as a miss. A summary not drawn from the page — text presented under a source chip for a page the product did not fetch — counts as not opened and is recorded in a footnote.

What the record promises. The ledger grows by one column a week and is never corrected: a week's cells are written once. When the conditions change — a page submitted to a search index, a product added or dropped, a page's text rewritten — the date goes on the record, so readings before and after stay separable. Every reading says on this date, this was observed. None says what a product does or does not do.

Limits, stated. Invented pages are not news pages, and reach on them depends on search indexes this project does not control. The grid is sixteen articles a week, read through free public interfaces with their quotas. A single week is a sample of a moving system; the value is in the series.

4The test pages

The fixed arm, as it stands. Word counts are the visible text, tags stripped. Nothing under /f/ reports real events, and the bylines are invented — disclosed in this site's robots.txt. The addresses below are given as text and are not linked from anywhere on this site, and Googlebot is asked not to crawl them: on 24 September 2026 it was observed that a page held in Google's index is answered from the index rather than fetched, and the answer is a refusal, while a page not held is fetched and read. The fixtures exist to be fetched. Every other crawler is welcome on them, and the ledger records each week whether any search index holds them regardless.

5Results

Everything measured is on one page, the ledger, which grows by a column each Monday from 28 September 2026:

• Fixed pages — which pages each product reached, week by week; and on the pages reached, which of the four fields the summary carried.
• Fresh grid — one row per week: how many of the sixteen summaries named the author, by product, with outlet, date and link in the footnote.
• Notes on the record — what the numbers do and do not claim, and the dated interventions the series is read against.