Sources: Search before Google II · Discovery
← Back to the Search before Google plate
Source. The chrome is the AltaVista homepage of 12 April 1997, held in reference/2-discovery/search-before-google/: two selects, size=55, maxlength=800, a Submit button, the <DL> promo block and the dot separated footer with Surprise sitting between Legal and FAQ. The masthead is Digital's real artwork, recovered from the 1997 captures and served from this origin rather than hotlinked: 7,324 bytes across five GIFs at 204, 59, 42, 65 and 35 pixels wide on one 73 pixel row, which is the 405 the page assembled. The tip line under the query box is quoted off the 11 May 1996 homepage, and it is the vendor telling you to do the thing this plate is about. There is no stylesheet in the dialect because the original had none: it was one <body> tag of color attributes, and the Legal panel shows it verbatim with vlink and alink missing their hash.
The scoring is the documented behavior and nothing else. From about_our_technology.htm: the NI2 indexer indexes each word from every page
and saves each instance of each word, including instances of different capitalization
. From addurl_meta.htm: the META description and keywords are indexed as words, the description is returned in place of an abstract, and the limit is 1,024 characters. From help_simple.htm: a word is any string of letters and digits
split on white space or punctuation, which is the tokenizer literally, plus the case rule, the trailing asterisk, the plus and the minus. Every one of those is implemented and nothing else is. The 1,024 sentence does not say whether the cap is each field or both; this reads it as both and prints the ambiguity beside the counter.
What could not be sourced is not invented, it is marked on the page in gray. Digital never published how it weighted those signals into an order. help_advanced_ranking.htm returns 404 in every capture the CDX index holds, and no results page for the engine was ever archived: all 431 archived cgi-bin/query?pg=q URLs are 404, 301, 302, 403 or 410. So this plate weights nothing. Position is the raw count of matching instances, ties held in index order, and the Detailed form prints the arithmetic. What the real Detailed form showed is unknown, so that control has been repurposed as the scoring X-ray and labeled. The results header and the rank numbers are the plate's own; 1997 results were not numbered.
Two rows of the corpus are real and fourteen are not. Pink Poodles Inc is AltaVista's own worked example out of addurl_meta.htm, verbatim down to http://pink.poodle.org/ - size 3k - 29 Feb 96, and the <head> in the editor opens as the exact three lines Digital printed. So the page you are told to optimize is the one the vendor's manual optimized first. Cool Site of the Day is a real page captured 26 December 1996, and only its two META fields are here: 300 characters, 29 terms, with csotd in it twice, on a page that was selling nothing. Its body text is not reproduced. The other fourteen pages, their titles, bodies, byte sizes and dates are written for this plate and quote nothing.
The ending is one site contradicting itself, and both halves are quoted. addurl_meta.htm: It is however possible for you to control how your page is indexed by using the META tag.
addurl.htm, the same site the same month: The usual technique is to submit pages with numerous keywords ... We strongly discourage the use of these techniques. AltaVista is an index, not a repository for pages of low or misleading informational value.
Nothing published between them says how many keywords is numerous. The brake was real and it was a person at Digital reading your page and deciding.
Checked live, 21 August 2026. Marginalia returns 200 and is byte identical to the capture taken three weeks earlier, all seven positioning bullets still on it, AGPL still stated, Explore still a peer of Search in the nav, and old-search.marginalia.nu still answering for text based browsers. Every masthead GIF was parsed for its own dimensions rather than trusted from a table. The two help pages were pulled from the Wayback Machine while this was being built, because the query rules had to come from somewhere and the folder did not have them.
Deliberately absent. Usenet search, which sat behind the other select and gets one line saying so. The Advanced Query syntax page and its boolean operators. The DoubleClick banner, which is a gray 468x60 with the ad server's path in it and no more. Any real network request. Multiple pages of results. Any weight on where in a document a word appeared, which the indexer demonstrably recorded and Digital never explained. And any claim about how many sites stuffed keywords or when the practice stopped, because the only period example in hand is one page from December 1996.