Google indexed the books. Why did nobody build Coffeetable?

The answer in one page

You are asking the correct question late, but not too late.

The uncomfortable answer is that most of Coffeetable was technically possible before modern LLMs. Full-text OCR and indexing were mature enough for Amazon to search 120,000 books in 2003. Passage retrieval existed as an academic field by 2002. Semantic retrieval existed decades before chat interfaces, and Google demonstrated a recognizably similar product in 2018: Talk to Books searched every sentence in more than 100,000 books for a conversational response to a user’s question. Google says it delighted millions of users, proved the technology, and closed in June 2023.^[1]

The thing that was not solved was permission plus economics.

There are four different products hiding inside “search books”:

1. Recommend a title using metadata, reviews, purchases, or text-derived features.
2. Search and navigate a book to say where an idea, name, or passage occurs.
3. Show a bounded sample so someone can decide whether to buy or borrow it.
4. Assemble and deliver a new reading made from verbatim pages of one or several books.

Publishers have repeatedly accepted the first three when they are controlled discovery tools. Amazon’s 2003 Search Inside the Book launched with 190 participating publishers and reported that sales growth for included titles exceeded the control group by 9% in its first five days.^[2] Google Books lets partners provide the complete file for indexing, restricts what users can see, disables printing and copying, reports preview and purchase-link activity, and sends readers to sellers.^[3]

The fourth product is where Coffeetable crosses the line from finding a book into making a reading product out of books. That can be a much better consumer experience. It is also the use most likely to replace a sample, an ebook sale, a library ebook licence, a subscription read, or a separately licensed anthology. Google’s court victory does not give a general right to do that. It protects Google Books as designed: full-text indexing plus deliberately small, discontinuous snippets that are difficult to reconstruct into meaningful reading. The Second Circuit explicitly relied on those limits.^[4]

That is why the Google fight matters. Google won the right to build an index; it did not win a blank cheque to build a new book-reading market. A proposed settlement had actually tried to create that market—with consumer access, institutional subscriptions, previews of up to 20%, a Book Rights Registry, and 63% of relevant revenue for rightsholders—but the court rejected it in 2011 as overbroad, especially because it would grant forward-looking, opt-out rights over absent and orphan-work owners.^[5] The legally safe remainder was discovery, not consumption.

The publisher proposition for Coffeetable is therefore not:

Let us use your books because the model is clever and discovery is good.

It is:

Give us a controlled runtime display licence for selected titles. We will show a bounded number of exact pages to a qualified reader at the moment of expressed intent, preserve author and edition attribution, send the reader to buy or borrow the book, report every page served, prevent corpus extraction and model training, and pay you on an agreed event.

That proposition is plausible, not proven. Publishers already licence excerpts, samples, ebooks, subscriptions, classroom extracts, search previews, and AI uses. Wiley reported $40 million of AI-content-licensing revenue in fiscal 2025.^[6] Kindle Unlimited pays from a pool according to pages first read.^[7] The industry is capable of inventing a licence when it sees money and control.

But Coffeetable has not yet proved either side of the marketplace:

• Consumer enthusiasm proves that the output can be useful. It does not prove willingness to pay, repeat use, book conversion, or that free excerpts grow rather than replace purchases.
• The public-domain product proves that the interaction can work. It does not prove that a publisher will license the in-copyright version.
• A publisher saying “interesting” proves nothing. The useful evidence is a written pilot permission covering real titles under workable display and payment terms.

My recommendation is a four-week rights-and-economics test, with product work frozen except what is required to run the test. Continue only if one credible publisher or rights-owning group permits at least 100 commercially relevant titles, or two smaller rightsholders sign equivalent pilots, and the permitted experience still feels like Coffeetable. If every serious conversation collapses into title-by-title permissions, author-by-author approvals, one-page previews, or economics that cannot support the product, move on or keep it deliberately public-domain. Do not spend another month making the unlicensed version better while hoping the rights problem later becomes somebody else’s job.

Confidence in the central conclusion: high. Confidence that a publisher pilot can be won: moderate, and intentionally unproven.

1. What Coffeetable actually is

The current implementation is more ambitious than “recommend a good book.” Its loop is:

1. accept a reader’s messy situation, memory, question, or desired feeling;
2. choose real books;
3. search their full text;
4. retrieve exact page labels and inspect the pages;
5. select the passages that answer the reader;
6. bind verbatim pages from one or, normally, several books into a focused Kindle-like reading;
7. let the reader continue in Coffeetable or export the generated EPUB to Kindle.

The code is explicit that the product’s point is the “conversation between books,” not a list of recommendations. The composed reading normally uses two to four books, one to three pages from each, arranged as an argument. The public client currently enforces a hard boundary to reviewed public-domain Project Gutenberg editions.

This produces a useful diagnostic distinction:

The mistake would be to negotiate all six as if they were “book discovery.” A publisher will not. Search metadata, displaying a page, creating a cross-title compilation, and distributing a durable EPUB invoke different rights, contracts, and perceived threats.

The sharpest statement of the company risk

The more magical Coffeetable becomes for the reader, the less it resembles Google’s protected snippet search.

That does not make the product impossible. It means the magical part is the part that must be licensed and paid for.

2. Yes, this could have been built before LLMs

What was already available

By 2003, the necessary mechanical stack existed:

• digitise print books;
• run OCR to recover machine-readable text;
• build an inverted full-text index;
• search inside millions of pages;
• rank passages by term relevance;
• extract names, places, products, themes, and other entities;
• infer similar books from content and behaviour;
• return a page around the match;
• measure the click and connect it to a purchase.

Semantic search is not an LLM invention. Latent semantic indexing was published in the late 1980s to address the mismatch between the words in a query and the words in a relevant document.^[8] Passage-retrieval systems were a recognised research area by 2002.^[9] Google had page-level similarity infrastructure across billions of scanned pages by 2011.^[10]

Modern LLMs materially improve three things:

1. Intent translation. “I am about to fire a friend” can become literary concepts, candidate works, and likely search language without the user knowing keywords.
2. Editorial selection. A model can inspect retrieved passages and arrange a coherent sequence across authors.
3. Interface. The user can ask in ordinary language and refine the result conversationally.

Those are real improvements. They make the old search index feel like a librarian rather than a database. But an LLM does not produce the underlying right to ingest, retrieve, display, or redistribute the words.

Google built the closest precursor

Google Research launched Talk to Books in 2018. A person typed a question or statement; a neural language model compared it with every sentence in more than 100,000 books and returned passages that looked like conversational responses. The result displayed the matching sentence with surrounding context and a link to the book.^[11]

Google’s own postmortem says:

• millions of people used it;
• semantic matching worked;
• the underlying techniques moved into products including Shopping, Maps, Gmail, and Hangouts;
• the experiment closed in June 2023 because it had “taught us all it has to teach.”^[1:1]

Google does not publicly say that publisher opposition or poor economics killed it. It would be false to invent that causal claim. What the public record does establish is narrower and more useful:

• the interaction existed;
• users liked it at meaningful scale;
• Google treated it as a research experiment, not a standalone book business;
• it returned bounded passages and links, not durable, stitched multi-book readings;
• the semantic technology was valuable in higher-frequency Google products even if the book interface was not strategic.

In other words, Talk to Books validates the behaviour while weakening any claim that the retrieval mechanic itself is a moat.

3. The Google Books fight, correctly understood

2004–2005: Google copies first and asks later

Google announced major library partnerships in December 2004. Participating libraries let Google scan books; Google retained scans and machine-readable text and gave digital copies back to libraries. For books supplied voluntarily through the Partner Program, publishers could authorise previews. For library books, Google scanned many in-copyright works without obtaining permission from every author or publisher.

The Authors Guild sued in September 2005. Five publishers, represented by the Association of American Publishers, sued in October 2005.^[12]

This was not fundamentally an argument about whether search is useful. It was an argument about who may copy a complete work, who sets the licence, who gets paid, and whether a technology company may create a new market by making rightsholders opt out after the fact.

2008–2011: the settlement tried to create the missing market

The parties proposed a $125 million class settlement. Its commercial architecture was remarkably close to the infrastructure a book-intelligence product still needs:

• a Book Rights Registry to identify rightsholders and distribute money;
• cash payments of at least $60 for eligible previously scanned principal works;
• consumer purchases of digital access;
• institutional subscriptions;
• public-library access terminals;
• previews of up to 20% for covered books;
• 63% of covered revenue allocated to rightsholders;
• possible later models including subscriptions, print-on-demand, custom publishing, downloads, summaries, and compilations, subject to the agreement.^[5:1][13]

This is the crucial historical fact: people did not merely try indexing. They tried to turn the index into a licensed commercial reading system.

Judge Denny Chin rejected the amended settlement in March 2011. The central problem was not that digital book access was undesirable. The settlement would resolve allegations about Google’s past scanning by granting sweeping rights over future uses, including works whose owners were absent from the litigation. It used an opt-out structure, raised orphan-work and international-rights problems, and could give Google an advantage that competitors could not reproduce. Chin said the arrangement went too far; he suggested an opt-in structure would address many objections.^[14]

That distinction matters for Coffeetable. A court rejected a universal, opt-out shortcut. It did not reject voluntary, opt-in licensing.

2012: publishers make a private peace; authors continue

Google and the publisher plaintiffs settled separately in 2012 on confidential terms. Public descriptions said publishers could choose whether digitised titles stayed in Google Books, were removed, or were made available for broader preview or sale. The deal restored publisher control rather than creating a universal compulsory market.^[15]

The Authors Guild continued litigating.

2013–2016: Google wins—but the design limits are the holding

The district court found Google’s use fair in 2013. The Second Circuit affirmed in October 2015; the Supreme Court denied review in April 2016.^[16]

The Second Circuit held that making complete internal copies was fair use for full-text search and the then-designed snippet view. Its reasoning depended on several constraints:

• the public did not receive the scans;
• a result showed only a tiny part of a page;
• snippets were divided mechanically, not selected as complete ideas;
• one page in ten was unavailable for snippet display;
• a user could not simply iterate queries to recover the book;
• the display helped determine whether the book was relevant without offering a meaningful substitute for reading it.^[4:1]

The court was unusually explicit: the snippets were designed to provide enough context to evaluate interest without revealing enough to threaten the author’s copyright interests.

The legally important sentence for Coffeetable is therefore:

Google won because it prevented search from becoming reading.

That is an inference from the opinion, not a quotation from it, and confidence is high.

HathiTrust confirms the same boundary

The Second Circuit held in 2014 that HathiTrust’s creation of a full-text searchable database was fair use. A general user could learn which books and pages contained a term but did not receive book text. The court also approved full-text access for certified print-disabled readers, a different purpose with a distinct public-interest analysis.^[17]

Again: full internal copy, yes; public consumptive display, generally no, absent permission or a separately justified use.

Internet Archive shows the other side of the line

The Internet Archive scanned print books and lent digital copies under a “controlled digital lending” theory. Even under a one-owned-to-one-loaned model, the Second Circuit held in September 2024 that this was not fair use. The digital copies served the same purpose as the originals, copied whole books, and competed with an existing library-ebook-licensing market.^[18] Internet Archive did not seek Supreme Court review, ending the case in December 2024.^[19]

The case did not decide the legality of short passages, search results, books without a currently sold ebook, or public-domain works. It nevertheless makes the market-substitution issue impossible to ignore: a well-intentioned digital access product loses when it supplies a replacement for a licensed book market.

The 2025–2026 AI fights make purpose limitation even more important

In 2025, a federal judge held in the Anthropic litigation that training on lawfully acquired books could be transformative while leaving Anthropic exposed for building a central library from pirated sources. A $1.5 billion settlement over roughly 500,000 works was ultimately approved in July 2026.^[20]

That result is not a licence for Coffeetable. Model training that aims to create different outputs is legally and economically different from retrieving the original pages for a reader.

More immediately, in July 2026 Hachette, Cengage, Elsevier, author Scott Turow, and his company filed a proposed class action alleging that Google used works obtained for Google Books, Google Play Books, and other limited purposes to train Gemini without permission.^[21] The claims are allegations, not adjudicated facts. But the publisher position is unmistakable: access granted for one product does not silently expand to another purpose.

Coffeetable’s licence must therefore state separately:

• whether text may be indexed;
• whether it may be embedded or otherwise used for retrieval;
• exactly how much may be displayed;
• whether excerpts from different books may be combined;
• whether a temporary reader may be generated;
• whether a downloadable EPUB may be created;
• whether any model training is allowed (the sensible pilot answer is no);
• territories, editions, duration, security, reporting, and deletion obligations.

4. Who tried, what survived, and what each attempt teaches

The pattern across all nine

Products converge on one of five stable positions:

1. Index but do not let the index become the book. Google and HathiTrust.
2. Show controlled samples to cause a transaction. Amazon and Google partner previews.
3. Display everything because the user or institution paid for licensed access. Kindle, O’Reilly, library ebook platforms.
4. Use only public-domain or explicitly open works. Project Gutenberg and Coffeetable today.
5. Display without permission and accept that the product is legally unstable. Shadow libraries and the defeated Internet Archive theory for in-copyright ebooks.

There is no sixth historical example where a company sustainably assembles arbitrary verbatim pages from copyrighted trade books into a portable reading without either a licence or a purchase entitlement.

5. Why Google did not turn its index into Coffeetable

1. Its legal victory depended on restraint

The easiest answer is the most important. The more complete and satisfying Google’s returned passages became, the weaker the exact reasoning that protected them. A search engine wanted the user to locate a source and leave for the source. A reading product wants the returned pages themselves to be the experience.

2. The unit economics were bad for a horizontal search company

Books are enormous relative to their transaction frequency. Scanning, OCR correction, metadata reconciliation, edition matching, rights processing, and secure hosting are expensive. A person may buy a handful of books a year. Microsoft stated the problem plainly when it closed its programme: it needed a sustainable model for the search engine, consumer, and content partner and preferred verticals with higher commercial intent.^[22:1]

Google could absorb the index as mission infrastructure. It had less reason to become the rights administrator for every passage rendered.

3. Rights are fragmented below the publisher logo

“Talk to publishers” is necessary but not always sufficient.

The copyright owner may be the author; the publisher holds only specified exclusive rights. The relevant rights may differ by print, ebook, audio, territory, language, edition, excerpt, anthology, educational use, and new digital use. Older agreements may not convey ebook rights at all. The Authors Guild notes that many pre-1995 contracts omitted electronic rights, and the RosettaBooks litigation demonstrated that a grant to publish “in book form” did not automatically block authors from separately licensing ebooks.^[27]

Penguin Random House’s own permissions workflow says:

• excerpts and portions require a permission request;
• an anthology is a recognised separate use;
• one request may be required per book;
• PRH may distribute a book while the originating publisher retains permission authority;
• some uses cannot be licensed without author approval.^[28]

For Coffeetable, a cross-book reading can look less like “preview these books” and more like “publish a custom digital anthology.” That phrase should be tested directly with rights departments rather than hidden behind AI terminology.

4. Discovery and substitution are hard to distinguish in advance

Publishers can observe a buy-link click. They cannot easily know whether the person who read six excellent pages from four books subsequently bought all four, bought one, or went to sleep satisfied and bought none.

This uncertainty encourages conservative limits. The product must produce the evidence publishers lack.

5. The large platforms already had a satisfactory local optimum

Amazon benefits from search, samples, Kindle purchases, Kindle Unlimited, and customer lock-in. Google benefits from search, partner previews, and Play sales. Apple benefits from store recommendation. None needs to open the corpus and economics to an independent cross-store reading layer.

A startup’s opening is not that incumbents could not build it. It is that their vertically integrated model gives them little incentive to create a neutral, publisher-controlled layer that works inside ChatGPT, Claude, the web, and Kindle.

That is a distribution thesis, not yet a moat.

6. How publishers are likely to think about Coffeetable

Publishers are not a single mind. Trade publishers, university presses, technical publishers, educational publishers, agents, estates, and independent authors control different rights and have different economics. Still, their likely decision frame is predictable.

What they can like

1. Backlist discovery. A relevant page can revive a title that no storefront is actively merchandising.
2. Better qualification. The reader has expressed a specific need and judged the author’s actual words before the purchase.
3. A defence against summaries. Coffeetable can make the source more visible instead of letting a chatbot paraphrase the book into invisibility.
4. Attribution. Exact author, title, edition, ISBN, and page context can travel with every passage.
5. Measurability. Page served, reading completion, save, buy, borrow, sample continuation, and Kindle-send events can be reported by title.
6. New revenue. A display fee, qualified-lead fee, revenue share, subscription pool, or minimum guarantee could be incremental.
7. Control. The publisher can select titles, exclude windows or pages, set territories, change limits, and withdraw a work.

What they will reasonably fear

1. Substitution. A beautiful ten-page reading may satisfy the user instead of creating a book sale.
2. Aggregation. Small excerpts from many books can create a commercially valuable anthology even when no single excerpt is long.
3. Purpose creep. Search access quietly becomes model training, summarisation, or a permanent corpus.
4. Loss of format and window control. A downloadable EPUB is not equivalent to an expiring web preview.
5. Author and agent conflict. The publisher may lack the relevant right or owe consultation, approval, or a share of subsidiary-rights revenue.
6. Territorial and edition errors. The US publisher may not control UK, Indian, translation, illustration, or quoted third-party material.
7. Decontextualisation. A page may be accurate but misleading when detached from the book’s argument or plot.
8. Security and leakage. A queryable corpus can be attacked systematically even if the normal UI shows little.
9. Bad precedent. A cheap startup licence can become the benchmark demanded by larger AI platforms.
10. Opaque economics. “Exposure” is not compensation, especially after two decades of platforms claiming that traffic would be enough.

Current licensing behaviour says “controlled yes,” not “categorical no”

The industry simultaneously fights unlicensed use and sells controlled licences:

• Wiley reported $40 million in AI-content-licensing revenue in fiscal 2025, up from $23 million in fiscal 2024.^[6:1]
• HarperCollins sought individual author permission for a limited licence involving selected books; the Authors Guild used the episode to argue that AI rights are often reserved to authors rather than silently owned by publishers.^[29]
• The Association of American Publishers told the US government in 2025 that voluntary licensing markets for high-quality publisher content already exist and should be encouraged.^[30]
• Copyright Clearance Center already operates collective permissions mechanisms for excerpts and digital reuse in other institutional contexts.^[31]
• Kindle Unlimited proves that page-consumption accounting can be an accepted royalty unit when participation and economics are explicit.^[7:1]

The conclusion is not “publishers love AI.” It is: publishers licence uses when the grant is explicit, the files are controlled, the value is measurable, and they retain bargaining power.

7. What winning the fight would actually look like

Do not begin by trying to win a legal doctrine. Begin by making infringement irrelevant through a licence narrow enough to sign and useful enough to test.

A. Split the system into an index right and a display right

The system should be able to answer “this book is relevant” without assuming it may show the book.

Each title needs machine-enforced fields such as:

• rights owner and contracting entity;
• controlled edition and ISBN;
• allowed territories and languages;
• index/search permission;
• embedding or retrieval permission;
• maximum visible words/pages and rolling percentage;
• whether adjacent pages may be shown;
• whether multiple titles may appear in one reading;
• whether the result may be cached and for how long;
• whether download/EPUB export is allowed;
• buy, borrow, or subscription destination;
• payment rule;
• exclusion windows and excluded pages;
• licence expiry, revocation, and deletion date.

The rendering service—not the model prompt—must enforce these rules.

B. Start with a portfolio pilot, not individual excerpt requests

A viable first proposal:

• 100–500 selected backlist titles;
• one publisher, imprint, distributor, agent collective, or rights-owning independent cohort;
• one territory and one language;
• 90 days;
• web/app display first;
• one to four pages per title in a reading, with a strict per-user rolling cap;
• maximum aggregate pages per title per user over the pilot;
• no full-corpus browsing or raw-text API;
• no model training;
• no permanent EPUB until explicitly approved;
• author, title, edition, and purchase/borrow action on every page;
• monthly title-level reporting;
• immediate revocation control;
• a modest minimum guarantee plus a variable success payment.

Starting with publisher-selected backlist solves several fears at once: low current discoverability, lower new-release cannibalisation, simpler windowing, and a credible claim of incremental demand.

C. Treat Kindle export as a separate licence

An expiring, access-controlled reader and a downloadable EPUB are not the same product. The EPUB creates a durable copy, can travel outside Coffeetable’s controls, and resembles a custom publication.

For the first in-copyright pilot, keep “Send to Kindle” to one of these forms only:

1. send a purchase/sample link to Kindle;
2. send a manifest that deep-links to already owned Kindle books;
3. create an EPUB only from public-domain passages;
4. create a copyrighted-passage EPUB only when the licence expressly authorises durable export.

Trying to smuggle export inside “preview” will destroy trust in the negotiation.

D. Make the publisher dashboard answer the substitution question

For each eligible title, report:

• qualifying user intents;
• appearances in candidate sets;
• appearances in rendered readings;
• pages and percentage served;
• completion and dwell time;
• saves and repeat opens;
• clicks to sample, buy, borrow, or publisher page;
• completed purchases when attribution is available;
• incremental conversion against a metadata-only or tiny-snippet control;
• user requests blocked by rights limits;
• security anomalies and attempted extraction.

The experiment needs a control group. Compare:

• title recommendation only;
• recommendation plus a tiny search snippet;
• recommendation plus a licensed page-first reading.

The business claim is supported only if the page treatment improves a publisher-valued outcome enough to offset any substitution.

E. Offer economics publishers already understand

Test three structures rather than theorising indefinitely:

1. Qualified-action fee: pay when a reader completes the excerpt and clicks buy/borrow/sample.
2. Attributed revenue share: share revenue from sales or subscriptions caused by Coffeetable.
3. Reading pool: allocate part of subscription revenue according to first-time qualified pages read, analogous in structure—not necessarily rate—to Kindle Unlimited.

A minimum guarantee may be necessary because the publisher is taking platform and precedent risk before conversion is proven.

F. Make author benefit explicit

Publishers do not necessarily own every new digital use, and authors will object if the product is presented as a publisher-only windfall. The proposal should specify:

• whether authors opt in;
• how author approval is collected when required;
• how revenue enters the existing subsidiary-rights split or a new split;
• whether authors can see title-level usage;
• how they can exclude passages or withdraw;
• that their text is not used to train a replacement-writing model.

8. The publisher conversations to have now

Do not ask, “Do you like the idea?” That invites politeness. Ask questions that force the rights and economic structure into the room.

Rights questions

1. For which titles do you control the right to display verbatim digital excerpts outside your own store?
2. Does combining excerpts from several titles create an anthology, compilation, enhanced-ebook, or other subsidiary-right use in your contracts?
3. Which uses require author or agent approval?
4. Are there differences for backlist, out-of-print, ebook-unavailable, illustrated, translated, or estate-controlled titles?
5. Which territories and editions could a pilot cover without additional clearance?
6. Can you grant portfolio-level rules, or must every excerpt be approved title by title?
7. Does a temporary hosted reader differ from a downloadable EPUB in your rights analysis?

Product and fear questions

8. At what page, word, or percentage limit does discovery become substitution for you?
9. What content must never be selected automatically: endings, poems, recipes, exercises, illustrations, tables, or sensitive chapters?
10. Would cross-title readings be acceptable if every title is clearly separated and linked to purchase?
11. What security, audit, retention, and deletion terms are non-negotiable?
12. What would make this feel like another platform extracting value rather than a new sales channel?

Economics and proof questions

13. Which outcome matters most: sample starts, sales, library borrows, backlist engagement, or direct licence revenue?
14. What reporting would allow you to judge incrementality?
15. Which payment model can your contracts and royalty systems process today?
16. What minimum guarantee would make a 90-day pilot worth the internal work?
17. Which 100 titles would you deliberately choose to test this, and why those titles?
18. What measured result would make you expand from 100 to 10,000 titles?

The closing question

If I send a two-page pilot term sheet next week, who besides you must approve it, and what is the first objection that person will raise?

That question distinguishes an interesting conversation from a sales process.

9. A four-week decision test

Week 1: package the exact offer

Produce three artifacts, no more:

1. a two-page publisher pilot term sheet;
2. a five-minute demo using only public-domain books but showing the proposed rights controls;
3. a one-page security and data-flow diagram showing ingestion, retrieval, display enforcement, reporting, and deletion.

Build no new consumer feature unless a live publisher conversation requires it.

Weeks 1–3: conduct 20 serious interviews

Target a mix:

• 5 trade-publisher rights, digital, strategy, or business-development leaders;
• 3 university or specialist presses;
• 3 literary agents or estates;
• 3 independent publishers or author collectives;
• 2 ebook distributors or aggregators;
• 2 collective licensing or permissions specialists;
• 2 authors with meaningful backlists and clean rights.

For every conversation, record:

• rights controlled;
• acceptable display boundary;
• author-approval requirement;
• allowed territory and format;
• acceptable economics;
• named decision-maker;
• next document and date;
• explicit reason for no.

Do not count compliments, introductions, or “circle back later” as progress.

Weeks 2–4: make three concrete pilot offers

Ask for a defined title list and written permission. The objective is not a strategic partnership announcement. It is the smallest enforceable grant that lets real users test in-copyright pages.

In parallel: measure the consumer side honestly

With the current public-domain catalogue, measure:

• request-to-reading completion;
• pages actually read;
• repeat use within 7 and 30 days;
• percentage who save or send to Kindle;
• percentage who ask to continue with a source book;
• willingness to pay for the experience without pretending the in-copyright catalogue already exists;
• whether readers prefer a stitched reading over three recommendations plus conventional samples.

Consumer praise is qualitative evidence. Return behaviour and payment are economic evidence.

10. Go, narrow, or stop

Continue toward a licensed company if all are true

1. One credible rights holder grants at least 100 commercially relevant titles, or two smaller rights holders grant an equivalent combined pilot.
2. The grant permits enough text and cross-title assembly that the result still feels materially better than recommendation plus retailer samples.
3. The rights holder can approve at portfolio level without unbounded title-by-title transaction cost.
4. A workable payment and author-accounting path exists.
5. Real readers return or pay, not merely admire the demo.
6. The licensed page treatment improves a publisher-valued outcome over the control.

Narrow to a public-domain/open catalogue if

• users return and pay for the interaction;
• publisher rights remain too slow or expensive;
• the public-domain corpus is sufficiently valuable in education, philosophy, history, religion, or classics;
• the product is positioned honestly as a new reading layer for open literature, not as the future universal library.

That can be a good product. It is a different-sized company thesis.

Stop or materially change direction if any two are true after the test

1. Serious rights holders uniformly require individual title and author clearance that cannot be aggregated.
2. Permitted limits reduce the output to something no better than existing samples.
3. Durable export—the behaviour users most value—is categorically unavailable.
4. Users like receiving free pages but neither return, pay, nor move toward the source books.
5. The licensed experience cannibalises publisher-valued actions rather than increasing them.
6. Minimum guarantees and reporting obligations make healthy gross margins implausible.
7. The only available path depends on a novel fair-use theory for consumptive display.

The seventh is a hard stop for the venture-backed version. Google spent a decade and enormous legal resources establishing a narrower doctrine than Coffeetable would need.

Final judgment

The idea is not invalidated by the fact that Google could have built it. Nearly every valuable product could have been built by an incumbent. The more damaging fact is that Google did build much of it, reached millions of users, and still did not turn it into an enduring book product. That kills the lazy thesis that semantic book retrieval is itself a new company.

The remaining thesis is sharper:

A neutral, licensed reading layer can use conversational intent to place exact author pages in front of a qualified reader, compensate the rightsholders, and create more book demand than it substitutes.

That is a real thesis. It is also entirely unproven.

The consumer side has shown enough life to justify one serious rights test. It has not shown enough to justify ignoring rights. The next founder work is not more retrieval code, a nicer reader, or a broader scraped catalogue. It is obtaining a real grant of rights on terms that preserve the experience, then measuring whether publishers make more money and readers come back.

If the licence cannot be won, the universal version of Coffeetable is not currently a company. If the licence can be won but readers do not return or pay, it is not a company. If both happen, the historical failure was not that nobody imagined the product; it was that nobody assembled permission, intent, display, attribution, payment, and distribution into one loop.

That is the fight worth testing.

Sources